Search

Objective To discuss the feasibility of directly processing paper-based ECG by using deep learning method, and to propose a method for identifying atrial fibrillation ( AF) from paper-based ECG rapidly and accurately. Methods The 12-lead ECG signals of 1 040 AF and 1 344 non-AF patients were selected from CPSC 2018 Challenge data, and were plotted as ECGs. These ECGs were printed out on papers, and rescanned into images. The paper-based ECGs with white background were obtained by a series of preprocessing. These paper-based ECGs were used to construct AF recognition dataset; the artificial intelligence algorithm (AIA) was trained, validated and tested using the real labels provided by the Challenge as the gold standard. Results The sensitivity, specificity and F1 score of the AIA based on paper-based ECGs are 0. 957, 0. 978 and 0. 969, respectively, and the area under receiver operating characteristic curve is 0. 994. The Grad-CAM based feature visualization results show that in paper-based ECGs, the position of P wave and F wave at the onset of AF provide main references for the algorithm to identify AF, which is consistent with the ECG diagnostic criteria of AF in clinical practice. Conclusion The AF recognition algorithm based on paper-based ECGs shows good performance. The visualization results verify the high efficiency and feasibility of directly analyzing paper-based ECGs by using AIA, which could provide guidance for clinical practice. [ABSTRACT FROM AUTHOR]

Biological small molecules such as microRNAs play an important role in biological processes and can positively or negatively regulate gene expression. Studying the relationship between microRNAs and genes is of great significance for the maintenance of homeostasis and the treatment of diseases. In this paper, the deep learning method is used to predict the relationship between microRNA and gene targeting, and the TransformerMGI model is proposed. In the feature engineering stage, aiming at the problem that it is difficult to extract the potential information of biological sequences accurately, the GP-GCN method based on graph convolutional neural network and DNA2Vec model are used to extract the potential information of microRNA and gene data, respectively, and the representation embedding matrix of microRNA and gene data is obtained. In terms of models, the TransformerMGI model introduces power normalization to improve the classical deep learning models. After feature extraction of microRNA and gene data in this paper, two representation matrices are obtained, which are respectively put into the TransformerMGI model. Through the internal Attention mechanism of the TransformerMGI model, the self-and mutual feature information of microRNA and gene are aggregated and correlated. Finally, the probability of microRNA regulating genes is predicted. In this paper, the area under the ROC curve and the exact recall curve are used as the performance evaluation indicators to evaluate the proposed model compared with other existing models. The experimental results show that the AUC and AUPRC scores of the TransformerMGI model proposed in this paper can reach more than 0.91, which is better than other existing models. The TransformerMGI model can only rely on the base sequence information of microRNA and gene without considering the biological principles and genomic background, and realize the prediction of microRNA target genes, which provides a useful deep learning method for the subsequent prediction of microRNA target genes. [ABSTRACT FROM AUTHOR]

The federated learning framework that uses autoencoder model to detect malicious model updates is an excellent defense framework for poisoning attacks. However, the existing autoencoder-based models face several challenges such as training difficulties and limited anomaly detection capability. In view of the above problems, this paper proposed an algorithm for detecting malicious model updates of federated learning based on β-VAE: the central server stabilized the training dataset by suppressing random attributes of the training samples, and used the dataset to train β-VAE anomaly detection model in real time. The model computed the anomaly score of task model updates uploaded by the clients, and then to detect and exclude malicious model updates based on the dynamic threshold of anomaly score. This paper evaluated the performance of the proposed algorithm on three federated learning tasks. Specifically, these tasks include classification on the MNIST dataset using the logistic regression (LR) model, classification on the FEMNIST dataset using the convolutional neural network (CNN) model, and character prediction on the Shakespeare dataset using the recurrent neural network (RNN) model. The experimental results show that, the task model under this algorithm achieves higher accuracy compared to other defense algorithms. This indicates that in Non-IID scenarios, the algorithm exhibits strong robustness against poisoning attacks in federated learning. [ABSTRACT FROM AUTHOR]

BACKGROUND: The development of artificial intelligence in the medical field is rapidly advancing, with increasing research on its applications in the field of bone trauma. Through bibliometric analysis, this paper analyzed the research hotspots of artificial intelligence in the field of bone trauma in recent years, and predicted the future research trend. OBJECTIVE: To summarize the development history, research status, hot spots, and future development trends of artificial intelligence technology in the field of bone trauma to provide new insights for future research. METHODS: This study selected relevant literature from the Web of Science core database, covering the period from the inception to August 2023, and retrieved 420 articles related to the application of artificial intelligence, machine learning, and deep learning in the field of bone trauma. After manual screening, 202 articles related to this article were exported, and Citespace software was used for visual analysis of cooperation of countries, institutions, cited journals, citation analysis, keyword co-occurrence, and other aspects. RESULTS AND CONCLUSION: (1) The overall number of publications from the 202 selected articles showed an upward trend, indicating significant research potential for future studies. The country with the highest centrality and the highest publication volume was the United States. The University of California (USA) was the most prolific research institution. (2) The top five most commonly used keywords in bone trauma research using artificial intelligence were deep learning, artificial intelligence, bone density, machine learning, and diagnosis. The keyword with the highest centrality was bone density, and the keyword with the highest frequency was deep learning. (3) The top 10 most cited reference papers provided comprehensive insights into the feasibility of applying artificial intelligence techniques to the diagnosis of bone trauma from various perspectives. Among them, eight papers focused on bone and joint injuries and deep convolutional neural networks. One paper discussed the use of deep learning in detecting osteoporosis in CT scans to prevent fragility fractures, while another paper explored the correlation between the application of artificial intelligence in identifying changes in skin texture and the recognition of bone characteristics. (4) In the future, the research hotspots of artificial intelligence will mainly focus on the specific study of fractures caused by bone and joint trauma and osteoporosis. The research trend mainly focuses on improving the performance of artificial intelligence algorithms, using new artificial intelligence technologies to accurately classify and quickly and efficiently diagnose bone injuries, especially for the diagnosis of complex and hidden fractures. By establishing finite element analysis models, more standardized evaluations of bone injuries can be achieved. [ABSTRACT FROM AUTHOR]

In order to solve the problems of difficulty in obtaining paired data sets and poor quality of enhanced images in the process of low-light image enhancement, the implementation of unpaired low-light image enhancement was studied by improving the cycle generative adversarial network model. In the generator part, a U-NET model integrated with the Vision Transformer structure was used to replace the original generator model, so as to improve the cycle consistency and content retention of image transformation, and effectively deal with the problem of long-distance spatial correlation commonly existing in image research. In the discriminator part, PatchGAN was selected to replace the traditional discriminator according to the characteristics of image research to improve the discrimination ability of image details. At the same time, the identity consistency loss function was introduced to improve the image quality. The results show that compared with the traditional method, the improved model in this paper has better subjective visual effect, and also has a corresponding improvement in the objective evaluation index. It can be seen that the improved model in this paper is effective. [ABSTRACT FROM AUTHOR]

To address the issue of the traditional fish hypoxia detection methods, which have low accuracy and require a lot of labor, a Prune-YOLOv5s-based hypoxia risk assessment method for farmed fish is proposed. This paper introduces a hypoxia risk assessment method for cultured fish based on the Prune - YOLOv5s algorithm. This method first collects data on aquatic surface respiration (ASR) performed by fish under hypoxic conditions to create a data set for fish hypoxia. The dataset is then utilized to train the YOLOv5s model. Then, the lightweight and improved YOLOv5s model was used to monitor the behavior of fish surface respiration during hypoxia in real-time. The introduction of the ASR coefficient allows for quantifying ASR instances in fish, which indicates hypoxia risk. The fish hypoxia assessment module is designed to evaluate the risk of hypoxia. The improved performance of the YOLOv5s model before and after modifications and the accuracy of the fish hypoxia assessment module are tested through the fish hypoxia experiment. The test results show that compared with the YOLOv5s model, the detection accuracy, model size, inference speed, and detection speed of the PruneYOLOv5s model have been significantly improved. Among them, the detection accuracy of the 65% Prune YOLOv5s model, which has the best comprehensive performance, has been increased by 0.6% compared with the original model. The size of the model is reduced to 45. 3% of the original model. The inference speed is improved by 23.8%, and the detection speed is enhanced by 31.4%. The fish hypoxia assessment method achieves 97.4% accuracy in the test set of 39 test videos and performs well in the hypoxia cycle experiment. The research indicates that the Prune-YOLOv5s-based hypoxia risk assessment method for cultured fish can effectively detect hypoxic conditions and provide accurate risk alerts, showing high feasibility for practical application. [ABSTRACT FROM AUTHOR]

Freezing damage in spring is one of the serious agricultural meteorological disasters for winter wheat production in the Huang-Huai-Hai plain. In order to effectively identify freezing damages of winter wheat in spring, this paper used a deep learning long short-term memory model LSTM to identify the occurrences of freezing damages of winter wheat in spring in Henan province, which was an important winter wheat planting area in China. The spatiotemporal distributions and decadal changes of freezing damages of winter wheat in spring based on the multi-source meteorological, crop, and disaster data from 99 meteorological stations from 1981 to 2020 were explored. The results indicated that: (1) the optimized LSTM model effectively identified freezing damages of winter wheat in spring. The daily minimum temperature simulated by LSTM model from 2017 to 2020 showed an average absolute percentage error (MAPE) of 8.73% and a goodness of fit (R2 ) with 0.90 compared to the actual minimum temperature. Based on the disaster data and freezing damage index for winter wheat in spring, the actual disaster situations from 2017 to 2020 were found to be consistent with the simulated results by the optimized LSTM model. (2) From 1981 to 2020, the harm of mild freezing damage to winter wheat in spring in Henan province had gradually weakened, and the high frequency area of mild freezing damage moved from the Northeast to the East. The overall distribution of mild freezing damage of winter wheat in spring in Henan province between 1981 and 2020 was higher in the northeastern Henan, and lower in the southwestern and southeastern Henan, with an average frequency of 0-1.75 times·10y-1 during the past 40 years. The frequency of mild freezing damage of winter wheat in spring decreased per decade, gradually decreasing from 0.843 times per decade to 0.157 times per decade. The high frequency area of mild freezing damage of winter wheat in spring moved from the Northeast to the East. (3) From 1981 to 2020, the frequency of severe freezing damage of winter wheat in spring in Henan province showed a trend of first increasing and then decreasing. The high frequency area of severe freezing damage of winter wheat in spring moved from the Northeast to the East. The overall distribution of severe freezing damage during the 40a period was higher in the East than in the West, and higher in the North than in the South. The average frequency of severe freezing damage in the 40a period was 0-2.75times·10y-1 . The frequency of severe freezing damage of winter wheat in spring increased from 0.508 times per decade to 0.857 times per decade and then decreased to 0.289 times per decade. The high frequency area of severe freezing damage moved from the Northeast to the East. The overall trend of severe freezing damage of winter wheat in spring in Henan Province was decreasing, but the frequency of extreme weather events under climate warming was increasing. Strengthening the researches on the impacts of major agricultural meteorological disasters on agricultural production in various regions in China under climate change was still one of the future research focuses. Various accuracy evaluation indicators of the deep learning LSTM model for identifying freezing damage of winter wheat in spring proposed in this study were improved, and the simulation results of freezing damage by LSTM model were basically consistent with the actual occurrences. Our results can provide ideas and methods for large-scale freezing damage identification of winter wheat under global climate change, and also have certain reference values for other agricultural meteorological disaster identification. [ABSTRACT FROM AUTHOR]

The estimation of multiple matching models between wide baseline or large angle images is a quite challenging task in image processing. The existing algorithms can be used to estimate multiple matching models and their inliers between images well, but their results are prone to matching pairs mis-classification issues. In order to accurately estimate the multiple matching models and allocate matching pairs, this paper proposed an image multi-matching model estimation algorithm based on the aggregation of matching points of the same model (AMPSM). Firstly, for improve the proportion of correct matching pairs, it filtered out incorrect matching pairs based on the distribution characteristics of correct matching points in the neighboring region. Furthermore, based on the different matching model degrees to which the matching pairs belong, searched for the suspected intersection matching pairs of multiple models, that was interference matching pairs. Meantime, for reducing the impact of interference matching pairs on the accuracy of matching classification, they were removed. Afterwards, for improve the clustering degree of matching points with the co-model, the position was dynamically moved based on the distance between the points within the same model and the center of gravity of the point set during the sampling process. Finally, classifying clustered matching points by Mean Shift to obtain a multi matching model. And the proposed method was compared with classical framework based algorithms RANSAC, PROSAC, MAGSAC++, GMS, AdaLAM, PEARL, MTC, Sequential RANSAC, and deep learning based algorithms SuperGlue, OANet, CLCNet, CONSAC, etc. Results indicate over 30% increase in the inlier rate, 8.39% reduction in the mis-classification rate of multi model estimation. It is concluded that the new algorithm has significant advantages in incorrect matches filtering and multi-model estimation. [ABSTRACT FROM AUTHOR]

Current researches on NDP flooding behavior detection mainly focus on detecting RA flooding and NS flooding behaviors, and there is insufficient flooding detection for other messages of the NDP protocol. Moreover, traditional threshold rule detection methods suffer from poor dynamics and low accuracy, while most of the Al-based detection methods can only perform binary classification detection, and there are still challenges in performing multi-classification detection. In short, there is a lack of corresponding research in multi-classification flooding detection of all messages of NDP protocol. Therefore, this paper proposed a multi-classification detection framework for NDP protocol flooding behaviors, and proposed a flooding behavior detection method for NDP protocol based on time interval characteristics. The framework constructed the first multiclassification dataset for NDP flooding detection through the processes of traffic collection and data processing, it compared and used 5 machine learning and 5 deep learning algorithms to train the detection model. The experimental results show that the detection accuracy of the XGBOOST algorithm in machine learning can reach 99.18%, and the detection accuracy of the Transformer algorithm in deep learning can reach 98.45%. Compared with the existing detection methods, the accuracy is higher. Meanwhile, the detection framework can detect 9 types of flooding behaviors for all 5 types of messages of NDP protocol and classify the flooding behaviors into multiple types. [ABSTRACT FROM AUTHOR]

Nowadays due to the dependence of fault diagnosis method based on deep learning on well-labeled training dataset, which will lead to the problem that deep neural network can easily overfit those noisy labels and affect the generalization of network under the condition of label noise. In order to achieve accurate recognition of equipment operating conditions in the network trained with label noise, this paper proposed a fault diagnosis method via contrastive learning and neighborhood sample analysis. Firstly, the method used contrastive learning to pre-train the model, which could reduce the embedding distance of similar samples in the feature space and achieved improving the ability of optimizing the feature representation ability of the network. Then, the method utilized the feature similarity to find each sample's closest neighbors to estimate the reliability of training labels which could separate all training samples into a clean or noisy subset and implemented label correction on noisy subset. After that, it established a more reliable training subset. Lastly, the proposed method made use of label reweighting and consistency regularization to enhance robustness of network. In particular, two networks got trained simultaneously where each network used the dataset division from the other network during the training process, which could mitigate confirmation bias caused by single network model training framework. The experimental results on public dataset demonstrate that proposed method can verify and correct the noisy labels impressively well and maintain great fault diagnosis performance under the condition of high-level noisy labels. [ABSTRACT FROM AUTHOR]

Deep semi-supervised clustering aims to achieve better clustering results using a small amount of supervised information. However, the amount of supervised information is often limited due to the expensive labelling cost. Therefore, with limited supervised information, it becomes crucial to select the most valuable supervisory information for clustering. To address the above problem, this paper proposed a deep active semi-supervised clustering model (DASCM) which designed an active learning method that was able to select marginal texts containing rich information and further generated high-value supervised information containing edge texts. The model used this supervised information to guide the clustering, thus improving the clustering performance. The experimental results on five real text datasets show that the clustering performance of DASCM is significantly improved. This result verifies that supervised information generated using active learning methods that cover marginal text is effective in improving clustering. [ABSTRACT FROM AUTHOR]

To address the challenge of constructing spatial and temporal dependencies in existing traffic flow prediction methods, this paper proposed a new method called spatial temporal position attention graph neural network (ST-PAGNN), which utilized spatiotemporal location attention. Firstly, the graph neural network contained a location attention mechanism, which could better capture the spatial dependence of traffic nodes in the urban road network. Then, it used a gated recurrent neural network with trend adaptive transformer (Trendformer) to capture the local and global information of the traffic flow sequence in the time dimension. Finally, it used the improved grid search optimization method to optimize the introduced para-meters of the model, obtaining the global optimal solution with high time efficiency. The experimental results show that in the dataset PEMS-BAY, the evaluation indexes RMSE, MAE and MAPE of the ST-PAGNN method are 1.37, 2.57, 2.67%, 1.55, 3.64, 3.37%, 1.97, 4. 37 and 4.43%, respectively, when the prediction step size is 15 min, 30 min and 60 min, respectively. In the dataset METR-LA, when the prediction step size is 15 min, 30 min and 60 min, the evaluation indexes RMSE, MAE and MAPE of the ST-PAGNN method are 2. 73, 5. 16, 7.13%, 2.99, 5.97, 7.86%, 3.53, 7. 16 and 9.96%, respectively. The results show that the proposed ST-PAGNN method is higher than the existing models in the evaluation indexes under different granularities, which illustrates the effectiveness and superiority of ST-PAGNN in solving traffic prediction problems. [ABSTRACT FROM AUTHOR]

[Objectives] The objective of this study was to address the problems of region non-closure, edge non-fitting and noise in the semantic segmentation of remote sensing images based on deep learning, and to further correct the recognition errors of semantic segmentation. [Methods] Taking Funan County of Anhui Province and Huai’an City of Jiangsu Province as the research sites, this paper built its own farmland plot dataset, introduced the idea of scale segmentation considering the multi-scale characteristics of images and the DESTIN (delineation by fusing spatial and temporal information) segmentation algorithm based on phenology, and proposed a semantic segmentation method for farmland plots based on multi-scale and DESTIN constraints. [Results] The field semantic segmentation based on the deep model under multi-scale and DESTIN-constrained effectively improved the problems of unenclosed region, non-fitting edges, noise and block-like blurring in the model, and corrected the misidentification area of the deep model semantic segmentation to a certain extent, and the IoU index reached 94.08% and 90.79% on the two test sets, respectively, and increased the semantic segmentation of the field by 1.65% and 2.32% compared with the remote sensing image of the deep model, respectively. The extraction area of the field in the study area was more complete and the accuracy was higher. [Conclusions] Multi-scale and DESTIN constraints further improved the semantic segmentation problem of fields, which was helpful to improve the accuracy of field recognition in high-resolution remote sensing images. [ABSTRACT FROM AUTHOR]

In recent years, the incidence and mortality rates of cancer have been rising steadily, with drug resistance becoming a major challenge in cancer treatment. Traditional monolithic treatment approaches have proven ineffective in addressing the heterogeneity of tumor cells and their multiple resistance mechanisms, leading to suboptimal therapeutic outcomes. Drug combination therapy, as a critical strategy, aims to enhance efficacy and delay the development of drug resistance through the synergistic action of multiple drugs. However, conventional methods for screening drug combinations are time-consuming and costly. With the accumulation of data and advances in computational methods, artificial intelligence, particularly deep learning, has demonstrated great potential in predicting synergistic drug combinations for cancer treatment. Artificial intelligence technologies allow researchers to efficiently predict whether drug combinations exhibit synergistic effects, reducing experimental costs and identifying novel potential synergistic combinations. Nevertheless, artificial intelligence models still face challenges such as poor interpretability, insufficient feature integration, and a lack of labeled data. This paper provided a comprehensive review of the advancements in artificial intelligence applications for predicting synergistic drug combinations in cancer therapy. First, it discussed the mechanisms of drug resistance and the challenges of combination therapy, highlighting the limitations of traditional drug combination screening methods. Then, it presented the advantages and disadvantages of various deep learning models used for predicting synergistic drug combinations, including feedforward neural networks, graph neural networks, autoencoders, visible neural networks, Transformer and their extended models. In response to the common issues in current deep learning models, this review proposed several solutions, such as leveraging multimodal data to enhance model generalization, employing transfer learning and multitask learning to address data scarcity, and designing more interpretable models to facilitate clinical application. In the future, the field of synergistic drug combination prediction is expected to benefit from the development of standardized synergy metrics, improvements in model interpretability, integration of multimodal data, and effective handling of data limitations, further advancing the transition of models from laboratory research to clinical practice, ultimately providing more effective solutions for cancer treatment. [ABSTRACT FROM AUTHOR]

Computer aided design of drug molecules is a molecular simulation method by constructing new chemical entities with certain potential pharmacological activity in computer or theory. In the past decade, the rapid development of high-throughput omics technologies provides good data support, and researches opportunity for the molecular design of biological and chemical drugs. Moreover, modern society has higher requirements for biopharmaceuticals, which generally needs the features of high efficiency, non-toxic or low toxicity and strong targeting. As more and more structures of proteins relating to drug-targets are resolved through experimental methods, drug molecular design methods based on protein receptors become more feasible and important. Drug molecular design methods based on protein receptors are generally analyzed based on the three-dimensional structures of proteins and ligands, which makes the discovery of lead compounds more rational, and more scientific design can be carried out. With the accumulation of relevant experimental data and the development of algorithms such as deep learning, more scientific drug molecular design can be carried out, which speeds up the development process of new drugs to some extent and is more conducive to exploring corresponding molecular mechanisms. In this paper, the common strategies of protein receptor based drug molecular design methods are systematically reviewed and prospected. [ABSTRACT FROM AUTHOR]

Freezing damage in spring is one of the serious agricultural meteorological disasters for winter wheat production in the Huang-Huai-Hai plain. In order to effectively identify freezing damages of winter wheat in spring, this paper used a deep learning long short-term memory model LSTM to identify the occurrences of freezing damages of winter wheat in spring in Henan province, which was an important winter wheat planting area in China. The spatiotemporal distributions and decadal changes of freezing damages of winter wheat in spring based on the multi-source meteorological, crop, and disaster data from 99 meteorological stations from 1981 to 2020 were explored. The results indicated that: (1) the optimized LSTM model effectively identified freezing damages of winter wheat in spring. The daily minimum temperature simulated by LSTM model from 2017 to 2020 showed an average absolute percentage error (MAPE) of 8.73% and a goodness of fit (R²) with 0.90 compared to the actual minimum temperature. Based on the disaster data and freezing damage index for winter wheat in spring, the actual disaster situations from 2017 to 2020 were found to be consistent with the simulated results by the optimized LSTM model. (2) From 1981 to 2020, the harm of mild freezing damage to winter wheat in spring in Henan province had gradually weakened, and the high frequency area of mild freezing damage moved from the Northeast to the East. The overall distribution of mild freezing damage of winter wheat in spring in Henan province between 1981 and 2020 was higher in the northeastern Henan, and lower in the southwestern and southeastern Henan, with an average frequency of 0-1.75 times·10y-1 during the past 40 years. The frequency of mild freezing damage of winter wheat in spring decreased per decade, gradually decreasing from 0.843 times per decade to 0.157 times per decade. The high frequency area of mild freezing damage of winter wheat in spring moved from the Northeast to the East. (3) From 1981 to 2020, the frequency of severe freezing damage of winter wheat in spring in Henan province showed a trend of first increasing and then decreasing. The high frequency area of severe freezing damage of winter wheat in spring moved from the Northeast to the East. The overall distribution of severe freezing damage during the 40a period was higher in the East than in the West, and higher in the North than in the South. The average frequency of severe freezing damage in the 40a period was 0-2.75times·10y-1. The frequency of severe freezing damage of winter wheat in spring increased from 0.508 times per decade to 0.857 times per decade and then decreased to 0.289 times per decade. The high frequency area of severe freezing damage moved from the Northeast to the East. The overall trend of severe freezing damage of winter wheat in spring in Henan Province was decreasing, but the frequency of extreme weather events under climate warming was increasing. Strengthening the researches on the impacts of major agricultural meteorological disasters on agricultural production in various regions in China under climate change was still one of the future research focuses. Various accuracy evaluation indicators of the deep learning LSTM model for identifying freezing damage of winter wheat in spring proposed in this study were improved, and the simulation results of freezing damage by LSTM model were basically consistent with the actual occurrences. Our results can provide ideas and methods for large-scale freezing damage identification of winter wheat under global climate change, and also have certain reference values for other agricultural meteorological disaster identification. [ABSTRACT FROM AUTHOR]

Objectives: With the rapid development of the Internet, social media has become an important information source of emergency events. However, there are a lot of duplication, errors and even malicious contents in social media, which need to be effectively classified to provide more accurate information for disaster emergency response. Methods: Deep learning has greatly improved the accuracy and efficiency of text classification. This paper takes earthquake disaster as an example, and builds a multi-label classification model based on bidirectional encoder representation from transformers (BERT) transfer learning. Over 50 000 posts about 5 earthquakes are collected as training samples from SINA Weibo, which is a very popular social media in China. Each sample is manually marked as one or more labels, such as hazards information, loss information, rescue information, public opinion information and useless information. Results: By fine-tune training, the classification accuracies of the proposed model on training dataset and test dataset reach 97% and 92%, respectively. The area under curve score of each label ranges from 0.952 to 0.998. Conclusions: The results prove that the multi-label classification using BERT transfer learning is of high reliability. The proposed model can be applied to the emergency management services for earthquake events, which is beneficial for the rapid disaster rescue and relief. [ABSTRACT FROM AUTHOR]

Bone and soft tissue tumors occur in the musculoskeletal system, and malignant bone tumors of bone and soft tissue account for 0.2% of all human malignant tumors, and if not diagnosed and treated in a timely manner, patients may be at risk of a poor prognosis. Image interpretation plays an increasingly important role in the diagnosis of bone and soft tissue tumors. Artificial intelligence (AI) can be applied in clinical treatment to integrate large amounts of multidimensional data, derive models, predict outcomes, and improve treatment decisions. Among these methods, deep learning is a widely employed technique in AI that predominantly utilizes convolutional neural networks (CNN). The network is implemented through repeated training of datasets and iterative parameter adjustments. Deep learning-based AI models have successfully been applied to various aspects of bone and soft tissue tumors, encompassing but not limiting in image segmentation, tumor detection, classification, grading and staging, chemotherapy effect evaluation, recurrence and prognosis prediction. This paper provides a comprehensive review of the principles and current state of AI in the medical image diagnosis and treatment of bone and soft tissue tumors. Additionally, it explores the present challenges and future prospects in this field. [ABSTRACT FROM AUTHOR]

Currently, human sleep staging methods based on electroencephalogram (EEG) signals show a trend towards single-channel and deep network models, however, single-channel information acquisition makes EEG lose the positional information of brain regions, and the features characterizing sleep stages in EEG tend to be sparse and thus difficult to extract, at the same time, the common problems of deep networks-the artificial setting of the model and its training hyperparameters make the training process blind and inefficient, and these problems lead to the low accuracy of automatic sleep staging methods. Therefore, this paper proposed to use the inter-layer feature reuse function of DenseNet to explore the sleep state information hidden in EEG signals, and improved the DenseNet model for the low-frequency characteristics of single-channel EEG signals in the frequency domain and the long-range dependence of single-channel EEG signals in the time domain, so as to achieve the fast and accurate sleep staging of the human body. In order to further improve the performance of DenseNet, it used a deep deterministic policy gradient (DDPG) algorithm to optimize and automatically adjust the key hyperparameters of DenseNet using the reinforcement learning idea during the network learning and training process. The experimental results show that the staging accuracy of the algorithm model on the Sleep-EDFx dataset reaches 89.23%, and the overall performance is better than other advanced staging algorithms in recent years, demonstrating good application prospects. [ABSTRACT FROM AUTHOR]

Addressing the issues of slow speed and poor performance in multi-object grasping detection in unstructured environments, a method of performing object detection before grasping detection is pro- posed. In object detection, to accelerate the network's running speed, this paper improved the YOLOv5 network by employing depth wise separable convolutions and coordinate attention mechanisms. For the grasping task.a single-stage grasping pose detection algorithm was designed. Firstly, considering the in- terference present in unstructured environments, RGB-D images were selected as the input data for the grasping network, and GG-CNN was chosen as the backbone network. Secondly, to enhance the feature extraction capabilities of the grasping network, the parallel use of different size convolutional kernels in the Inception-ResNet module was utilized to broaden the network's receptive field. Additionally, the inte gration of a parameter-free three-dimensional attention mechanism enabled the network to focus more on grasping information features and suppress background noise. Finally, a grasping quality evaluation was employed to refine the grasping boxes, and the grasping box with the highest confidence score was out- put. The experimental results indicate that the improved object detection network has a parameter count of 2 776 708 and achieves 102 frames per second (FPS). On the public Cornell dataset, the improved grasping detection network achieves an accuracy of 96. 57% with a FPS of 54, 17. The combination of the two improved networks can be deployed on robotic arms and effectively accomplish grasping tasks in multi-object scenarios, making them suitable for practical industrial applications. [ABSTRACT FROM AUTHOR]

Objectives：In mid- to late- April 2024, an extreme heavy rainfall event occurred in Shaoguan City, Guangdong Province, inducing a large number of landslides in Jiangwan Town, Shaoguan. People lost connection with the outside world for nearly 36 hours, which aroused widespread social concern. Rapidly and accurately identifying the basic characteristics of landslides, development and distribution patterns and formation conditions is crucial for disaster emergency decision-making and risk elimination and disposal. Methods: Using the post-disaster optical remote sensing images and combining with deep learning model, the rainfall-induced landslides in Jiangwan Town, Shaoguan, were quickly and automatically identified. Results:After manually calibration, a total of 1 192 landslides were deciphered, with a total area of about 3.14 km². The scale of the landslides was dominated by small and medium-sized landslides, which were mainly distributed as an aggregated belt along the river in the northeast-southwest direction, with a significant characteristic of concentrated occurrence. Spatial statistical analysis showed that the landslides were mainly distributed on concave slopes with slopes of 10°-30° in the range of 200-300 m elevation. Further quantitative analysis of the geomorphic controlling factors of landslides using the random forest model and SHAP theory reveals that different topographic and geomorphic factors have different degrees of nonlinear effects on landslide formation, and that multiple factors such as elevation, slope, and catchment conditions are coupled to jointly control the formation of landslides. Conclusions: This paper highlights the great advantage of deep learning-based intelligent identification and analysis technology in the emergency investigation and formation conditions analysis of landslide disasters, which can provide important technical support for the rapid assessment of disaster losses and risk identification. [ABSTRACT FROM AUTHOR]

Objectives: Flood disaster is a very destructive natural disaster. The main reasons for its generation include heavy rainfall, storm surge and dam break. When flood disaster occurs in populated areas such as cities and towns, the flood will directly threaten the safety of life and property of residents. Also it will cause paralysis of land and underground transportation, interruption of water and electricity transportation. In the process of flood rescue, quick and accurate identification of flooded roads is conducive to planning appropriate personnel transfer and material transportation routes, and reducing subsequent losses caused by floods. Aiming at the problem that roads in the flood disaster scenario cannot be automatically and correctly identified, this paper proposes an end-to-end flooded road detection method based on deep learning. Methods: The proposed method uses a three-branch encoder-decoder structure and uses strip convolution, in which the efficient extraction of linear features is realized. And the coordination dual attention mechanism can effectively guide the network, and realize the identification of road areas. The method can effectively utilize the historical optical remote sensing images before the disaster and the real-time optical remote sensing during the disaster. The image is applied to detect the flooded and non-flooded roads in the disaster-stricken area, and a comparative experiment is carried out on the self-built dataset. Results: The experimental results show that the precison and recall rate are 0.838 1 and 0.666 8 on pre-disaster road, 0.796 6 and 0.607 4 on post-disaster road, 0.780 0 and 0.661 4 on affected road respectively. Conclusions: The proposed method has achieved the goal of automatically identifying the flooded roads in the flood-stricken area. The ability of detecting flooded roads and non-flooded roads can provide strong support for flood disaster rescue and reduce losses of life and property caused by flood disasters. [ABSTRACT FROM AUTHOR]

Re-identification technology for pedestrians and vehicles has been successfully applied in the field of intelligence analysis. However, there is a lack of research on re-identification technology for ship targets. In this paper, we propose a double-feature fusion-based maritime defogging re-identification network for intelligence analysis and supervision of ship targets. To reduce the impact of negative samples on features, we adopt a perspective-assisted adaptive query expansion method and a similarity-based feature fusion method. Furthermore, a defogging branch is embedded in the shallow layer of the re-identification branch. This branch utilizes weight sharing technology to extract fog-free features. The defogged image is then reconstructed using upsampling technology and the pyramid model, enhancing the recognition ability of the re-identification network in low-visibility scenarios. Finally, a pseudo-IOU based non-maximum suppression method is proposed to enhance the detection accuracy of ship targets. This method modifies the confidence of the detection frame. Experimental results demonstrate that the proposed method outperforms existing methods, and each module contributes to the network's performance. [ABSTRACT FROM AUTHOR]

Sericulture has a pivotal position in the agricultural production of China. The number of silkworms is one of the key indicators in the process of silkworm breeding, which is crucial to the feeding and healthy growth of silkworms. Accurate detection of the number of silkworms can ensure the quality of sericulture products. However, manual counting has not fully met the needs of large-scale production in recent years. There is a high demand to combine AI technology for silkworm counting. In this study, silkworm detection and counting were studied from the perspective of deep learning. Several challenges were solved in the detection of silkworms. Since the silkworms belonged to the small targets, it was difficult to detect the fewer features; When the number of silkworms was high, serious masking occurred to cause interference in the detection of densely distributed silkworms; It was difficult to ensure that the clear images of silkworms were captured from the actual production. The blur images were easy to deteriorate the accuracy of detection. Two tasks were covered as follows. 1) The image dataset of silkworms was constructed; 2) MAM-YOLO, a silkworm detection and counting model was proposed using multi-dimensional perception. The pictures of silkworms were first taken from the Guangxi Sericulture Technology Promotion Station. These pictures of silkworms were preliminarily processed. The polygon annotation was then selected, according to the body and distribution of silkworms. The pictures of silkworms were annotated with the open-source tool (Labelme). Finally, the dataset was expanded using image transformation and noise addition. The silkworm picture dataset was obtained for the subsequent training, verification, and testing. An improved YOLOv8 (MAM-YOLO) was proposed to detect the silkworm, in order to realize the accurate detection and counting. A silkworm detection and counting model was proposed using space-to-depth transformation and image deblurring. The fine-grained detection performance of blurred images was improved in the MAMYOLO model. The space-to-depth transformed convolution was utilized instead of stride convolution for the pooling layer in the feature extraction. The fine-grained features were extracted from the images, in order to further improve the detection of the low-resolution and small targets. In addition, the impact of blurred images on the detection performance was reduced by introducing a deblurring module into the model. RCS-OSA module was added to the YOLOv8 model. The network feature extraction was improved to integrate the information of different sizes in the receptive fields. The detection head of the original network was improved into the multi-dimensional sensing head. The three-dimensional features of size, space, and channel were aggregated to enhance the detection of silkworm targets. The loss function was replaced with the EIOU loss function. The regression accuracy of the detection frame was improved to reduce the object occlusion. Experimental results show that the MAM-YOLO more accurately detected the silkworm targets. The detection speed of mAP@0.5:0.95 reached 74.8% and 22.2 frames per second, indicating the better functions of silkworm detection and counting with high precision and high rate. The detection requirements were fully met in the actual work of the Guangxi sericulture station. The results show that the silkworm detection and counting system of the technical indicators met the application requirements. The finding can provide the detection services for the actual production work of the silkworm breeding station. In the end, a generalization experiment was conducted on the Global Wheat Challenge 2021 dataset. It was found that the MAM-YOLO was also achieved in the better detection performance of wheat ears, particularly on the small targets similar to dense occlusion.Sericulture has a pivotal position in China's agricultural production. The number of silkworms is a key indicator in the process of silkworm breeding, which is crucial to the feeding and healthy growth of silkworms, and accurate detection of the number of silkworms can ensure the quality of sericulture products. The traditional way of counting silkworms manually is inefficient, and it needs to be combined with technology to intelligently complete this kind of heavy work. In this thesis, from the perspective of deep learning, a silkworm detection and counting method is studied, and several problems in silkworm detection are solved: silkworms belong to small targets, which have fewer features so that they are not easy to detect; silkworms are densely distributed, and when the number of silkworms is high, serious masking phenomenon occurs which causes interference in the detection; it is difficult to ensure that the images of silkworms obtained in the actual production are clear, and blur images will make the detection affected. The main research content of this thesis is as follows: This chapter completed two main tasks: 1) Completed the production of silkworm image data set; 2) MAM-YOLO, a silkworm detection and counting model based on multi-dimensional perception, was proposed.In this paper, pictures of silkworms were taken by workers from Guangxi Sericulture Technology Promotion Station. First, these pictures of silkworms were preliminically processed. Then, according to the characteristics of the body and distribution of silkworms, the polygon annotation method was selected, and the pictures of silkworms were annotated with the open-source tool Labelme. Finally, the data set is expanded by means of image transformation and noise addition, and finally the silkworm picture data set is obtained for subsequent training, verification and testing.This chapter proposes an improved silkworm detection method based on YOLOv8, MAM-YOLO, which realizes accurate detection and counting of silkworm. Secondly, aiming at the problems of loss of fine-grained information and poor detection performance of blurred images in the MAM-YOLO model, a silkworm detection and counting model based on spaceto-depth transformation and image deblurring is proposed. The model utilizes space-to-depth transformed convolution instead of stride convolution and pooling layer in the feature extraction process, so that fine-grained features in the image can be extracted, which further improves the detection capability for low-resolution and small targets. In addition, the impact of blurred images on the detection performance is reduced by introducing a deblurring module into the model. By adding RCSOSA module to YOLOv8 model, the capability of network feature extraction is improved, and the information of different size receptive fields is integrated. The detection head of the original network is improved into multi-dimensional sensing head, and the three-dimensional characteristics of size, space and channel are aggregated to enhance the detection ability of silkworm targets. By replacing the loss function with EIOU loss function, the regression accuracy of detection frame is improved and the problem caused by object occlusion is reduced. Experiments show that MAM-YOLO can more accurately detect silkworm targets, and the detection speed of mAP@0.5:0.95 reaches 74.8% and 22.2 frames per second, which realizes the functions of silkworm detection and counting with high precision and high rate, and can meet the detection requirements in the actual work of Guangxi sericulture station. results show that the silkworm detection and counting system of the technical indicators have met the application requirements, and can provide detection services for the actual production work of silkworm breeding station.In the end, a generalization experiment was conducted on the Global Wheat Challenge 2021 data set, and it was found that MAM-YOLO also showed a high improvement in the detection performance of wheat ears, proving that the proposed method can generalize to other small targets similar to dense occlusion. [ABSTRACT FROM AUTHOR]

NeRF is a deep learning model aimed at modeling three-dimensional implicit spaces, and it holds significant value in the representation and rendering of 3D scenes. However, due to the complex training process, substantial computational resources, and time requirements, the usability and practicality of the NeRF algorithm are somewhat limited. Addressing the pain points of NeRF optimization has become a hot topic in the field of computer vision. This paper aimed to provide a comprehensive review of the optimization and application of NeRF. Firstly, it delved into the basic principles of NeRF and outlined the current optimization status from the perspectives of rendering quality, computational complexity, and pose. Secondly, it enumerated the application scenarios of NeRF to provide references for future, more efficient and practical algorithmic optimizations. Finally, it summarized the strengths and limitations of NeRF and proposed potential future directions tailored to harness the tremendous potential of NeRF in 3D rendering, scene synthesis, and beyond. [ABSTRACT FROM AUTHOR]

Camera pose estimation plays a crucial role in tasks such as autonomous driving and robotics, elucidating the direction and position of the camera in relation to a given scene through the estimation of its positional coordinates and angular deviations around the three coordinate axes. To facilitate the understanding of researchers in the realm of camera pose estimation, this paper comprehensively reviewed the current research status and latest progress in this field will. Firstly, it introduced the fundamental principles, evaluation indicators, and pertinent datasets associated with camera pose estimation. Subsequently, the review elaborated and summarized the two-stage model structure method and single-channel model structure method from the two key technologies of scene relationship construction and camera pose calculation. It conducted classification and analysis based on the diverse core algorithms and scene information employed, with performance comparisons drawn from indoor and outdoor public datasets. Lastly, it expounded the current challenges in the field and future development trends. [ABSTRACT FROM AUTHOR]

In recent years, there has been rapid development in intelligent aquaculture and fisheries resource conservation, leading to an increased demand for fish tracking technologies. Traditional fish tracking methods rely heavily on visual observation and tag tracking, which suffer from low efficiency, limited applicability, and low accuracy, hindering their widespread adoption. With the rapid advancement of deep learning in computer vision, deep learning-based fish-tracking technologies can provide accurate, objective, scalable, and automated tracking methods. Firstly, this paper introduces the tracking objects and four deep learning-based fish tracking methods:semantic segmentation, instance segmentation, object detection, and object classification. Secondly, it describes how fish tracking technologies capture fish trajectories, postures, fish quantities, and fish lengths, which are important tracking information for fish targets. Furthermore, the application of deep learning-based fish tracking technologies in fish diseases, fish feeding behavior, and fish health status is discussed. The paper also explores the main challenges of current deep learning-based fish tracking technologies, including low contrast and texture blurring, image color distortion, occlusion, and deformation, along with some corresponding solutions. Finally, the paper concludes and provides an outlook on the future development of deep learningbased fish-tracking technologies. It suggests that deep learning-based fish tracking technologies offer higher accuracy and objectivity, providing more solutions for practical applications in different scenarios. This technology is expected to play a more significant role in aquaculture management, fish scientific research, and marine environment conservation, offering more data and support to relevant fields. [ABSTRACT FROM AUTHOR]

In industrial production, defects are treated differently based on their severity, so it is necessary to classify defects. However, in actual production, the classification accuracy is often insufficient due to the presence of few easily confused defects, which requires conservative treatment of all defects in production practice, resulting in significant human and economic costs. To solve this problem, this paper proposes a method for interactive separating easily confused defects. This method separates few easily confused defects from other defects, ensuring that the classification results of the remaining majority of defects can be directly used. This method selects easily confused defects from the training data as one or more new defect categories, called virtual defects, so that the trained network can distinguish between virtual defects and other defects. This paper designs a visual interface to assist users in interactively selecting easily confused defects to construct virtual categories. CMOS defect data from actual industrial sites are adopted for effectiveness verification, and the results show that the proposed method can quickly classify few confusing defects and ensure that the classification accuracy of remaining defects meets the requirements of industrial applications. [ABSTRACT FROM AUTHOR]

[Objectives]Because of the limit of camera field, the walnut point cloud suffers from missing points, which detrimentally affects the accuracy of walnut phenotype analysis. To enhance the accuracy and expedite the efficiency of walnut grading and classification, this paper proposed a walnut external parameter estimation approach with three-dimensional reconstruction. [Methods]We introduced a coarse-grained realistic structure with the point cloud generation network(CGRSNet)to form a point cloud completion network with an encoding-decoding mechanism. Firstly, a new sampling algorithm was implemented to extract feature information of the original point cloud in different dimensions and thereby increased the feature extraction effectiveness of walnut phenotype. Afterward, the autoencoder was utilized to reconstruct the point cloud in multiple dimensions, and a complete walnut shape was eventually obtained. In the end, the parameters of the model were applied to estimate the volume of walnut. The walnut geometry model generated by multiple parameters was digitized into the point cloud data to feed CGRSNet and get a pre-trained walnut geometry model through shape completion. A fine-tuning process was then conducted with the real walnut point cloud data. [Results]The chamfer distance of the completion network for walnut data set was 0.196 cm. According to the linear regression analysis, the RMSE and R² score of the volume estimation were 2.47cm and 0.94, respectively, which were 0.88 cm and 0.01 higher than those of PF-Net, respectively. [Conclusions]The walnut external parameter estimation algorithm based on CGRSNet proposed in this paper presents satisfactory practicability and accuracy. [ABSTRACT FROM AUTHOR]

In recent years, with the rapid development and expansion of the global aquaculture industry, and the continuous enlargement of aquaculture farms, the industrialization, intelligence, and informatization of aquaculture have become a trend in the industry. China has become the largest producer of fisheries and aquaculture. Fish farming is an important component of aquaculture, and fish farming monitoring has become an important technology to enhance the efficiency, production, and management of fish farming. Fish farming monitoring can provide real-time and accurate data for farms, assisting farm managers in making decisions to improve efficiency and production. With the emergence of artificial intelligence technology in recent years, deep learning has rapidly developed and been widely applied in various fields such as image and audio recognition, natural language processing, robotics, bioinformatics, chemistry, and finance. The monitoring of fish farming focuses on the quantity, growth, behavior, and health status of fish. Using deep learning technology, we can quickly and accurately obtain information related to fish farming and enhance its efficiency and management. This paper presents a deep learning-based method for fish farming monitoring and reviews the literature progress in fish length measurement, fish counting, fish feeding, fish swimming behavior, and fish disease diagnosis. Although deep learning-based fish length measurement has achieved high accuracy in underwater environments, some errors still exist. The counting methods based on deep learning can be categorized into segmentation counting, detection counting, tracking counting, and density regression counting. Deep learning models based on video data have higher accuracy in recognizing fish feeding behavior than imagebased models. There have been many studies on fish tracking, but practical applications still face challenges such as fish feature extraction, the influence of fish size and obstructions, and occlusion issues. In fish disease diagnosis, it is necessary to establish standardized and shared fish disease datasets and utilize data fusion, data level information fusion, feature level information fusion, and decision level information fusion. This article also summarizes the main problems of deep learning-based visual technologies in fish farming monitoring from the aspects of monitoring data acquisition and transmission, dataset standardization and processing, deep learning model design, and the lack of business application in fish farming intelligent monitoring. The problems in data acquisition include a limited variety of experimental subjects, a small number of samples, and poor performance of experimental equipment. In the data transmission process, there are challenges in data security and realtime transmission. In terms of datasets, there is a low level of standardization and a lack of large-scale unified datasets. There is also a lack of research on large models and embedded models in deep learning model design. Furthermore, there is a realistic problem of insufficient business application in practical settings. The paper also proposes future research directions, including establishing fish farming monitoring datasets, super-scale parameter models for fish farming, edge computing for terminal monitoring devices, and digital twinning in fish farming monitoring, aiming to provide scientific references for the widespread application of deep learning in fish farming monitoring. [ABSTRACT FROM AUTHOR]

Objectives: The quality of network structure is the key to obtain excellent classification performance. Most of existing deep models are designed by the trial-and-error method. When the same model processes different hyperspectral images (HSIs), different network hyperparameters need to be set artificially to adapt to the data characteristics of different images, greatly increasing the difficulty of network design. Aiming at the problem of deep learning model design in hyperspectral image classification, this paper proposes an automatic network structure search method for HSI classification. Methods: In the proposed method, the differentiable structure search technology is first used to search the network structure on the source HSI data sets, then the deep network model is constructed by stacking cells, and finally the classification performance is evaluated on the target HSIs. It should be pointed out that, the proposed method only conducts one search on the source HSI data sets, and the obtained deep model can be applied to other target HSI classification tasks, which can effectively improve the utilization rate of the model. To improve the generalization ability and classification accuracy of the model, multi-source and multi-resolution HSIs are employed to construct the precollected data set, and the partial channel connection operation is introduced to further improve the search efficiency. Results: Three HSIs inculding Houston, Chikusei and Pavia Center are selected as source data. Four HSIs including University of Pavia, Indian Pines, Salinas and Houston 2018 are selected as the target HSIs. The source data is used to determine the network model and the target HSIs are used to evaluate the classification performance. Experimental results show that the proposed method can automatically search out a universal deep network model suitable for HSI classification task, and this model can achieve better classification performance than the conventional deep learning models. The overall classification accuracies of University of Pavia, Indian Pines, Salinas and Houston 2018 are 98.15%, 98.74%, 97.30% and 74.47%, respectively. Conclusions: Compared with other deep models, the model in this paper is automatically determined by differentiable architecture search algorithm according to the data characteristics of HSIs, effectively improving the classification accuracy of target HSIs while avoiding complex network design. Extensive experiments verify the applicability and effectiveness of the proposed method in HSI classification. [ABSTRACT FROM AUTHOR]

Since the objects in remote sensing images have the characteristics of arbitrary direction, dense distribution and large scale differences, object detection in remote sensing images has become a challenging problem. Aiming at this problem, this paper systematically reviewed the work related to rotating object detection in deep learning remote sensing images in recent three years. Firstly, this paper introduced the representation method and characteristics of rotated bounding box. Then, it analyzed the existing methods of rotating object detection in remote sensing images from four aspects: the feature extraction network, the generation of rotating anchor and candidate bounding box, the label allocation and sampling strategy, and the loss function. Next, it introduced the commonly used remote sensing image datasets for rotated object detection, and compared the performance of different algorithms. Finally, this paper prospected rotating object detection in remote sensing images. [ABSTRACT FROM AUTHOR]

Objectives In recent years, many deep learning-based classification models have been proposed. Because the size of histopathology images is extremely large, current methods usually divide it into many small patches of the same size and then build the classification model. The models first extract the features of patches and then use pooling methods etc. to integrate patch-level features into slide-level features for classification, in which the extraction of patch-level features and independent with the integration and classification process, making it impossible to adapt the feature extractor with the feedback of classification results. Therefore, they can not obtain the feature which is most useful for the whole slide image (WSI) classification. The objective of the paper is to solve the problems faced by above methods.Methods We propose an end-to-end attention pooling method based on a patch-scoring model. First, we build a patch scoring model to score each patch with the multiple instance learning methods, selecting patches based on the scores. Second, we utilize an attention-pooling module to integrate the features of selected patches to build a classification model. Besides, according to the character of the histopathology image, we propose to use a customized squared average function to normalize the attention weights to focus more on the high scored patches in positive WSI, making the classification results more interpretable.Results On the public available dataset CAMELYON16, the proposed method achieves F1 score at 0.644. And it achieves F1 score at 0.593 at the public available BACH dataset. When the new normalization function is used, the average attention weights difference between positive patches and negative patches is 0.017, which is much higher than when softmax function was used.Conclusions We can conclude from the experiments that the end-to-end attention pooling method is a better method for breast histopathology image classification. The proposed normalization function can make the attention module more interpretable. [ABSTRACT FROM AUTHOR]

Image stitching is an important branch in computer vision and computer graphics, and has a wide range of applications in 3D imaging and other aspects. Compared with the traditional image stitching framework based on feature point detection, the image stitching framework based on deep learning has stronger scene generalization performance. Although there are many research results on image stitching based on deep learning, there is still a lack of comprehensive analysis and summary of the corresponding research. In order to facilitate the subsequent work in this field, this paper sorted out the representative results in this field in the past 10 years. Based on the comparison between traditional stitching methods and deep learning-based image stitching methods, it collated and analysed the learning strategy and model architecture design, classical model review, and dataset from the three sub-problems of homography estimation, image stitching, and image rectangling in the research field of image stitching. It summarized some features of deep learning-based image stitching research methods and summarized the current research status in the field, and prospected the future research prospects. [ABSTRACT FROM AUTHOR]

For reducing the computational complexity of the high efficiency video coding (HEVC) standard, this paper proposes a complexity reduction approach based on deep learning. Firstly, a large-scale database for the coding unit (CU) partition is established, providing fundamental data for algorithm design and neural network training. Then, a CU partition map model is designed adaptive to HEVC, as an efficient representation of the CU partition across multiple adjacent CUs. Based on the model, a DenseNet-based hierarchical convolutional neural network structure is proposed to accurately predict the three-level CU partition in HEVC. The experimental results show that the proposed approach can save 57% of the encoding time on average while maintaining the coding efficiency, considerably alleviating the bottleneck of over-high encoding complexity in HEVC. [ABSTRACT FROM AUTHOR]

Predictive business process monitoring(PBPM) represents a vital research field within BPM that aims to accurately predict future behavioral events. At present, deep learning methods are widely used in PBPM research. However, most of these methods consider only a single event-control flow perspective and do not fuse the attribute-data flow perspective for process prediction. To address this issue, this paper proposed a method called the fusion multi-perspective(FMP) framework based on a two-layer BERT neural network. Firstly, the first layer of BERT was used to learn attribute-data flow information. Subsequently, the second layer of BERT learnt event-behavior control flow information. Finally, the FMP framework combined data flow and control flow to achieve multi-perspective process prediction. Experimental results on real event logs demonstrate that, compared to other research methods, the FPM framework yields higher accuracy in predicting the next event activity. This validates that the FPM framework, which merges multi-perspective views of processes, enables a more comprehensive and in-depth analysis of complex process behaviors while enhancing predictive performance. [ABSTRACT FROM AUTHOR]

Knowledge distillation algorithms have a great effect on the streamlining of deep neural networks. The current feature-based knowledge distillation algorithms either focus on a single part for improvement and ignore other beneficial parts, or provides effective guidance for the part that a small model should focus on, which makes the distillation effect insufficient. In order to make full use of the beneficial information of the large model and process it to improve the knowledge conversion rate of the small model, this paper proposed a new distillation algorithm. Firstly, it used the conditional probability distribution to fit the feature spatial distribution of the large model's middle layer, and then extracted the spatial attention maps that tended to be similar after fitting together with other beneficial information. Finally, it used the small convolutional layer, narrowed the gap between models, transmitted the transformed information to the small model to achieve distillation. Experimental results show that the algorithm has the applicability of multiple teacher-student combinations and the generality of multiple data sets, and compared with the current more advanced distillation algorithms, the performance is improved by about 1.19% and the time is shortened by 0.16 h. It has important engineering significance and wide application prospects for large networks' optimization and the application of deep learning on low-resource devices. [ABSTRACT FROM AUTHOR]

Voice conversion is a research topic in the fields of speech and artificial intelligence. The goal of voice conversion is to change the timbre of speech while preserving the content of the source speech, making it sounds like spoken by the target speaker. It is essential to ensure both the quality and naturalness of the converted speech. Voice conversion based on nonparallel data gains much attention currently, where models are trained using non-parallel multilingual speaker datasets, enabling many-to-many and any-to-any voice conversions. This paper provides a comprehensive summary and analysis of recent developments in non-parallel voice conversion. Firstly, we outline the early voice conversion techniques based on parallel corpus and their limitations. Then, we introduce and compare various approaches to voice conversion based on nonparallel data, providing a thorough analysis. Finally, a summary and outlook on voice conversion technology is provided. [ABSTRACT FROM AUTHOR]

Objectives: Recognition of building group patterns is an important part of the building generalization. Efficient methods for the recognition of building group patterns can effectively improve the quality of automated map generalization. Traditional recognition methods mainly include geometric methods and traditional machine learning methods, which are limited by the complex rule definitions and huge feature engineering. The graph convolution neural network (GCN) can overcome the limitations of traditional methods to some extent and has been successfully applied to the pattern analysis of buildings. However, there are few methods to recognize multiple building group patterns using GCN, and the existing GCN methods are almost based on spectral graph convolution, which do not consider local spatial information fully. Methods: As the spatial GCN is more efficient than the spectral GCN, this paper introduces a spatial GCN method DGCNN to recognize three building group patterns, including linear pattern, grid pattern and irregular pattern. The first step is to cluster the buildings into different groups, then some indices are selected to construct the feature vector of each group according to the Gestalt principles. Second, Delaunay triangulation (DT) and minimum spanning tree (MST) are chosen to construct graph structures of building groups. Finally, graph structures are used as input for the DGCNN model, and building group patterns are obtained after training. Results: The experiment selects the urban areas in Shanghai and we compare the accuracies with traditional spectral GCN method. The results show that the accuracy for the DGCNN model in the train set can reach 97.06%, and 95% for the test set. And the accuracies for the spectral GCN model in the train set is 90.90%, and 86.05% in the test set. Conclusions: The proposed method overcomes the limitations of spectral GCN, and it does not need graph Fourier transform. It can improve the recognition accuracy significantly and faces less difficulty in explaining the process of feature extraction. Hence it is an effective method for the recognition of building group patterns. [ABSTRACT FROM AUTHOR]

To address the low segmentation accuracy of multifidus muscle lesion sites in MRI images of patients with lumbar disc herniation, this paper proposes a new model to improve the U2-Net network with the goal that the encoding and decoding subnetworks are interconnected by a series of nested jump paths. To reduce the semantic missing of feature maps in the encoding and decoding subnetworks, the jump connections in the middle of RSU-7,RSU-6,RSU-5,and RSU-4 in the U2-Net model are redesigned, while the RSU-4F part remains unchanged. In addition, the channel attention module is added to enable the net to focus on channels of higher contribution to task, thus extract high quality multifractal muscle features. The experiments on the multifidus muscle MRI image dataset show that the improved U2-Net outperforms U-Net,U2-Net and U-Net++ network in indicators of Dice, HD and MIoU. It can be concluded that the proposed approach has good performance on MRI image segmentation of multifidus muscle, which can assist doctors to make diagnosis. [ABSTRACT FROM AUTHOR]

Network intrusion detection systems are crucial for maintaining network security, and there is currently limited research on intrusion detection scenarios with only a few abnormal markers of network data. This paper designs a weakly-supervised learning intrusion detection model, called WIDSAPL, based on the anomaly retention of data. The detection model consists of four parts: data transformation layer, representation learning layer, transformation classification layer, and anomaly discrimination layer. By using a set of learnable encoders to map samples to different regions and compress them into a hypersphere, the label information of abnormal samples is used to learn the classification boundaries of normal and abnormal samples, and the abnormal score of the samples is obtained. Testing the WIDS-APL system on four datasets demonstrates the effectiveness and robustness of the system, with improvements in the AUC-ROC values of 4.80%, 5.96%, 1.58%, and 1.73% respectively compared to other mainstream methods. Furthermore, there are enhancements of 15.03%, 2.95%, 4.71%, and 9.23% in AUC-PR performance. [ABSTRACT FROM AUTHOR]

[Objectives]This paper aimed to solve the problem of low target detection accuracy of apples with different ripeness under natural environment. [Methods]An improved YOLOv5s model SODSTR-YOLOv5s(YOLOv5s with small detection layer and omni-dimensional dynamic convolution and swin transformer block)was proposed for different ripeness apple detection. Firstly, the multi-scale target detection layer of YOLOv5s was improved, and the detection head of 160×160 feature map was constructed in Prediction to improve the detection accuracy of small-size apples with different ripeness. Secondly, Swin Transformer Block was integrated into the Backbone structure to strengthen the texture feature fusion of apples at the same level of ripeness, weaken the negative impact of the difference in texture feature distribution, and improve the generalization ability of the model. Finally, the Conv module of Neck structure was replaced by the dynamic convolution module ODConv, and the local feature mapping was refined to realize the full extraction of local apple fine-grained features. Experiments were carried out based on different ripeness apple data sets to verify the performance of the improved model. [Results]The precision, recall and average precision of the improved model SODSTR-YOLOv5s were 89.1%, 95.5% and 93.6%, respectively. The average precision of high, medium and low ripeness apples was 94.1%, 93.1% and 93.7%, respectively. The average detection time was 16 ms and the number of parameters was 7.34 M. Compared with the YOLOv5s model, the precision, recall and average precision of the improved model SODSTR-YOLOv5s increased by 3.8%, 5.0% and 2.9%, respectively, and the number of parameters and average detection time increased by 0.32 M and 5 ms, respectively. [Conclusions]The improved model SODSTR-YOLOv5s improved the detection ability of apples with different ripeness under natural environment. It can better meet the detection requirements of actual apple picking. [ABSTRACT FROM AUTHOR]

Implementing fast and accurate hand detection on embedded devices mainly face two challenges. Firstly, it is difficult for complex deep learning networks to achieve real-time hand detection. Secondly, the complexity of the scene leads to a decrease in the accuracy of hand detection algorithms based on RGB color images. Unlike mainstream RGB image based detection technologies, this paper adopted a lightweight hand detection algorithm based on ToF infrared images to attain precise and swift hand detection within the infrared images. Firstly, this paper gathered 22 419 static infrared images using this self-engineered equipment, thereby establishing an infrared dataset tailored for hand detection. Subsequently, it enhanced a general object detection algorithm to create a lightweight hand detection network known as Retina Hand, using two different lightweight networks, MobileNetV1 and Shuffle NetV2, as the backbone network of the model. Furthermore, this paper proposed an attention-enhanced feature pyramid structure called Attention-FPN. This structure integrated attention mechanisms to enhance the detection process. Ultimately, this paper conducted comparative experiments on the infrared dataset against conventional methods to validate the effectiveness of the method. [ABSTRACT FROM AUTHOR]

Traditional geological stratification methods usually rely on manual interpretation and empirical judgment, which have shortcomings such as cumbersome processes, heavy workload, and large impact from human factors. This paper aims to study the ideas and methods of regional engineering geological stratification based on machine learning, and convert the stratigraphic stratification problem into a sequence-to-sequence prediction task of geological body spatial units and a geological attribute feature classification task, so as to improve the accuracy of geological stratification. and efficiency. Based on in-depth research on the original data of engineering geology, this paper proposes to use deep learning methods to train and predict geological stratification of static cone test data, and use classification algorithms to stratify borrow hole soil samples and quantify them. The applicability of the relevant algorithms was evaluated. By comparing the results of traditional methods and machine learning methods, the advantages of machine learning in geological stratification are verified. This study provides a new geological layering idea for regional engineering geology research, which has important practical significance. [ABSTRACT FROM AUTHOR]

The intermittent, random and uncontrollable power generation characteristics of renewable energy pose challenges for the full utilization of green energy. The high energy consumption feature of the virtual power plant data center makes it an efficient absorption and regulation strategy for the intermittent (non-dispatchable) power in renewable energy. This paper proposes a method to predict the "source-load" dual-state of the virtual power plant using a long short-term memory network that incorporates time-embedded encoding. The results indicate that using the model presented in this paper can achieve proactive alerts for "power shortages" at 15-minute intervals, creating ample buffer time windows for container suspension and backup. Combined with container technology, it realizes dynamic energy consumption management in data centers, thereby enhancing the robustness of the virtual power plant data center against power supply-demand imbalances. This technology is of great significance for stabilizing grid operations, accelerating the application of green clean energy, constructing a service pattern for the energy ecosystem and speeding up the digital transformation of the grid. [ABSTRACT FROM AUTHOR]

The management of safety hazards in water conservancy engineering construction is transitioning towards informatization and intelligence. In order to efficiently mine valuable potential information from a large amount of unstructured construction safety hazard data, an intelligent recommendation system for construction safety hazard rectification based on deep learning is proposed. This paper is based on the TF-IDF algorithm to extract feature words of hidden danger, construct a safety hazard association Sankey diagram and display the information flow characteristics among construction sections, hazard features and hazard types. Then, this paper mines association rules in historical data based on the FP-Growth algorithm. In addition, the process of case retrieval recommendation is optimized by combining the sequence similarity matching algorithm and the Doc2Vec model. This paper uses 80 953 construction safety hazard information as the data source, which is recorded in the water resources allocation project of Pearl River Delta from 2019 to 2023. Example verification shows that the proposed method can match accurate rectification measures for current construction safety hazards, effectively assisting construction safety managers to identify and address hidden danger. [ABSTRACT FROM AUTHOR]

HD maps can provide accurate road information and rich semantic information, so that the automatic driving system can guide the vehicle to drive correctly. HD maps usually rely on manual labeling. The existing automatic labeling methods have low recognition accuracy in autonomous driving scenarios, resulting in low efficiency of HD map labeling. In order to solve this problem, this paper proposed a new semantic segmentation method for automatic labeling of HD maps called MapFormer. This paper proposed a multi-level feature fusion module, which enabled the model to aggregate different levels of detail and semantic information. Additionally, it proposed a novel boundary-decoupled joint decoder to improve the model's ability to handle intercategory boundaries. Experiments on the bird's-eye view dataset show that the proposed model not only achieves excellent performance in segmentation accuracy, but also has a clearer handling of category boundaries. The mIoU of MapFormer is 55.82%, which is higher than the SegFormer of 1.03%. The proposed method can improve the HD map labeling efficiency and labeling automation rate. [ABSTRACT FROM AUTHOR]

To address the problem of insufficient recommendation performance in session recommendation, this paper proposed a disentangled graph neural network model (GDST-GNN) using global graph diffusion and spatio-temporal awareness. Specifically, the model firstly constructed a global collaborative graph based on all sessions from a global perspective, and then employed graph diffusion as the message propagation paradigm for global representation learning of items to capture global information be- yond the current session. For the representation learning of the current session, this paper designed a disentangled spatio-temporal gated network modeling the complex transition pattern and temporal dependency pattern of items in the session, and then fused the learned global and local representations factor by factor. In addition, this paper employed a self-supervised task to achieve performance enhancement of the model. Finally, it generated session representations through attention networks to achieve accurate recommendation of items. Extensive experiments on four real-world datasets validate the effectiveness of the proposed model. [ABSTRACT FROM AUTHOR]

Multi-epoch interferometric synthetic aperture radar (InSAR) is a highly effective technique for monitoring deformation in urban areas. However, interpreting InSAR deformation can be challenging due to various factors, including inherent geometric imaging distortion, the intricate structure and deformation properties of targets in urban scenes, and the multiple scattering of microwave signals between objects in urban scenes. This paper discusses the challenges involved in interpreting time-series InSAR deformation: (1) Precisely identifying the location of deformation signals and linking them to their corresponding objects, i. e., determining where the deformation signal occurs, (2) understanding the mechanisms and factors that cause the detected deformation signals, i.e., determining what the deformation signal represents, (3) establishing the connection among the detected deformation signals, the deformation events, and the scattering mechanisms. We suggest a parametric framework to improve the accurate interpretation of InSAR deformation. This framework includes several factors, including kinematic characteristics (deformation rate, cumulative deformation, deformation gradient, and deformation model), geometric parameters (position, size, structure, orientation, and roughness), semantic information (land cover type, terrain morphology, texture, and auxiliary information on natural and anthropogenic disturbance) and physical properties (scattering mechanism, penetrability, extensibility, conductivity, and thermal conductivity). Our approach aims to enhance the representation of coherent points for a better understanding of InSAR deformation. This paper offers a comprehensive overview of the advancements achieved in extracting parameters of InSAR coherent points and interpreting deformation based on geometric parameters, semantic information, and physical properties. High-precision 3D positioning is crucial for InSAR fine monitoring in urban areas. It helps determine the source of deformation signals and facilitates the analysis of deformation mechanisms. Semantic information, such as 3D models, high-resolution optical images, laser point cloud data, and land use data, can aid in interpreting InSAR deformation. By combining InSAR deformation data with a deep learning approach, there is an opportunity to interpret deformations effectively. In urban environments, the scattering mechanism of ground objects is complex. Multiple scattering signals can provide effective observations of deformation and information about the target's size. However, combining the scattering mechanism of synthetic aperture radar signals to carry out parameter inversion and deformation mechanism interpretation of urban target terrain remains a challenge. The framework, which considers the geometric parameters, semantic information, and physical attributes of InSAR coherent points, will be crucial for deformation interpretation and mechanism cognition. This framework will enable fine deformation monitoring, intelligent recognition, and application in future urban scenes. [ABSTRACT FROM AUTHOR]

Objectives: The Mw 7.8 earthquake occurred in Turkey on February 6th, 2023, which is a rare powerful earthquake in recent years. Building collapse detection is an important part of post-earthquake damage assessment and emergency rescue. Facing the emergency rescue and post-earthquake disaster assessment, this paper aims to comprehensively utilize the intelligent interpretation technology of satellite remote sensing, coupled with high-resolution optical daily and night-time light remote sensing images, and to efficiently assess the collapsed buildings in the extreme disaster areas with low time and labor costs. Methods: In this paper, we present a multi-level rapid localization scheme for disaster-affected regions. First, the proposed approach involves the utilization of low- and medium-resolution night-time light data to estimate the extent of extreme disaster areas. Subsequently, high-resolution optical remote sensing imagery is employed to extract the collapsed buildings. By training a semi-supervised deep learning network with a small number of manually collected samples, the scheme automatically identifies collapse areas, thereby facilitating the precise localization of collapsed buildings. A high-resolution architecture network, HRNet, is employed in this study. HRNet has the advantages in both sematic and spatial feature representation, and can capture the collapsed buildings with muti-scale characteristics. Results: Six cities along the border between the Arab Plate and the Eurasian Plate are severely affected, from Adelman to Kahraman Marrash, and then to Antakia. A total of 2 377 collapsed buildings are detected in the extreme disaster areas. Except for Nurdaji and Antakia cities, where the proportion of collapsed buildings exceeded 2%, the proportion of collapsed buildings in the other four cities is close to 1%. After manual verification, the recall rate of the intelligent interpretation plan is between 74% and 93%. Conclusions: The efficiency and effectiveness of this approach are demonstrated through its application in assessing the building collapse situation after the Mw 7.8 earthquake in Turkey. The proposed work can serve as a timely reference for macro decision-making in post-earthquake emergency rescue operations. [ABSTRACT FROM AUTHOR]