Search

In order to improve the training speed and certain classification accuracy of CNN in the image classification, this paper started from BN, adjusted the affine transform of BN by adding new parameters, and proposed an improved BN, called BNalpha. Except fo r the neural networks with some specific structures, compared with the original BN, BNalpha could improve the training speed and certain classification accuracy of general neural networks without increasing the computational complexity. By analyzing and comparing the parameters of BN affine transform, this paper tried to explain the operation mechanism of BN partially, and conducted contrast experiment which covering different periods during training to illustrate that BNalpha was superior to the original BN. Based on the CIFAR-10 and CIFAR-100 datasets, by using various types of CNN structures, it compared and analyzed BNalpha and BN, and the experimental results verify that BNalpha can further improve the training speed and certain classification accuracy. [ABSTRACT FROM AUTHOR]

In order to improve the long-term prediction accuracy of PM2. 5, this paper proposed a TSMN (time series memory network) prediction model based on deep learning by taking air pollutants and meteorological factors as the influencing factors. The model consisted of two components. The local memory component used external memory to improve the long-range memory capacity of the model, and cooperated with the neighborhood component that modeled the multi-site spatial relationship to complete long-term prediction of PM2. 5 from the perspective of time and space. This paper compared the TSMN model with multiple models by using different evaluation indicators. Comparing the TSMN model with the better-performing CNN-1.STM model, the RMSE and MAE of this model goes down 5. 2% and 5. 7% respectively,and the R2 goes up 7. 5% . The experimental results show that the TSMN model can effectively improve the long-term prediction accuracy of PM2. 5 concentration. [ABSTRACT FROM AUTHOR]

The community detection method can mine hidden cluster structure in the network, which is of great significance for the analysis of the structure and function of complex network. In recent years, with the explosive growth of network data and the diversity of network evolution, a large number of methods and frameworks for community detection that can handle different scenarios have emerged. In order to deeply understand the research status and development trend, this paper reviewed the community detection methods. This paper firstly classified these algorithms, introduced each kind of algorithm in detail and made analysis and comparison, and then introduced some commonly used evaluation indicators and expounded the application scenarios. Finally, this paper looked forward to the future research directions of this field. [ABSTRACT FROM AUTHOR]

This paper constructed a new TerraSAR-X dataset named GDUT-Nansha and then proposed a new lightweight semantic segmentation algorithm using synthetic aperture radar images（ SAR）. It is difficult to apply traditional deep learning models to SAR datasets with small volume due to the tremendous amount of parameters. A revision scheme is, therefore, put forward to deal with such problem. This paper accomplished an improved lightweight convolutional neural network model named revised weighted loss ENet（ RWL-ENet） for SAR images. The current study introduced a weighted loss function to solve the problem of imbalance of training datasets. Compared with other classical convolution neural network models, the efficiency and robustness of the new dataset were validated. Meanwhile, RWL-ENet model attained 0. 884,0. 804, and 0. 645 in terms of three quantitative metrics, including pixel accuracy（ PA）, mean pixel accuracy（ m PA）, and mean intersection over union（ m IoU）. In addition, the parameters of this new proposed model are much less than other classic network models. [ABSTRACT FROM AUTHOR]

Aiming at the traditional model parallel methods for accelerating convolution neural network (CNN) tend to have low utilization, this paper proposed PipeCNN, which accelerated CNN with software pipeline. Firstly, this paper studied the forward propagation and back propagation, and then explored data correlation during training. Secondly, it parallelized CNN with the support of software pipeline, and then analyzed two feasible gradient updating methods in PipeCNN. Finally, it used circular queue to realize communication between two layers and then proposed a task allocation algorithm to divide CNN into working parts. Experiments show that the method can obtain good speedup and utilization while ensuring the accuracy of the model. It shows that software pipeline can effectively solve the problem of low utilization in model parallel, and accelerate the training of CNN. [ABSTRACT FROM AUTHOR]

As an important part of software, code comments play an important role in software maintenance, software re use . Automatic code comments generation technology aims to reduce the workload of manual comment, thus improving the efficiency of software development. Existing comments automatic generation methods can be categorized into rule-based, text summarybased, data-based, topic model-based, deep learning-based and other levels . This paper summarized the concept of automatic code comments generation, compared and summarized the various types of code comment automatic generation methods, sorted out and reviewed the research progress of code comment automatic generation in recent years. Finally, this paper analyzed the challenges of code comments automatic generation and look ed forward to the future research in this area. [ABSTRACT FROM AUTHOR]

Recent years, with the widespread use of video surveillance technology, video anomaly detection, which can intelligently analyze massive videos and quickly discover the abnormalities, has received wide attention. This paper aims to give a comprehensive survey on deep learning based video anomaly detection methods. Firstly, a brief introduction of video anomaly detection is given, including the basic concepts, basic tasks, modeling process, learning paradigms as well as the evaluation perspectives. Secondly, the video anomaly detection methods are classified into four categories: reconstruction-based, prediction-based, classification-based, and regression-based. Their basic modeling ideas, typical algorithms, advantages, and disadvantages are discussed in detail. On this basis, the commonly used single-scene video anomaly detection public datasets and evaluation indicators are introduced, and the performance of representative anomaly detection algorithms is compared and analyzed. Finally, summary is conducted, and the future development directions related to datasets, algorithm and evaluation criteria of video anomaly detection are proposed. [ABSTRACT FROM AUTHOR]

To deal with challenges of limited and small-size defects in product quality inspection, this paper proposed a method for surface-defect-detection of small-size with few training samples (Smooth-DETR). This method utilized DETR-based encoder-decoder to predict the classification and location of defects, which reduced the parameters and complexity. DETR has a strong global feature learning capability, which could obtain rich texture features of product surfaces with few samples, so that it is easy to detect defects that break the continuity of texture. The combination of Smooth-L1 loss and GIoU loss improves the regression accuracy on small-size defect samples. Experimental results show that the proposed method performs better than the existing state-of-the-art methods. Moreover, the average detection precision of the proposed method for 11 different classes of surface defects is higher than 98% with few training samples. [ABSTRACT FROM AUTHOR]

Aiming at the poor classification effect caused by complex background, large intra-class difference and high interclass similarity in remote sensing scene images, this paper proposed an attentional mechanism and residual contraction unit algorithm based on supervised contrast learning. Firstly, the algorithm improved the effective channel attention mechanism (ECA), and optimized the extraction of image features to be recognized. Then, the algorithm proposed a cooperative residual shrinkage unit algorithm, which is used to filter and eliminate redundant information of images. In addition, supervised contrast learning algorithm was used to enhance the generalization ability of the algorithm. Finally, experiments were carried out with remote sensing image data set and compared with the latest algorithms such as enhanced attention algorithm and scale attention mechanism algorithm. Experimental results show that the proposed algorithm improves the classification accuracy by 1.75% and 2.5% in AID Data Set with 20% training ratio. [ABSTRACT FROM AUTHOR]

In view of the shortcomings of traditional salient target detection methods in detecting multiple salient targets at different scales. This paper presents a salient object detection algorithm with deep multiplexing of multi-scale features. The network model consists of vertically stacked bidirectional dense feature aggregation modules and horizontally stacked multiresolution semantic complementary modules. First, the bidirectional dense feature aggregation module extracts semantic features of different resolutions based on the ResNet backbone network, and then performs adaptive fusion on the top-down and bottom-up paths in turn to obtain multi-scale representation features at different levels; The multi-resolution semantic complementation module fuses the multi-scale features of two adjacent levels to eliminate the mutual crosstalk between features at different levels and enhance the consistency of prediction results. The experimental results on 5 benchmark datasets show that the method can achieve the highest Fmax, Sm, MAE of 0.939, 0.921, 0.028, and the detection rate can reach 74.6 fps, which has better detection performance compared with other comparison algorithms. [ABSTRACT FROM AUTHOR]

The traditional research on educational knowledge graph is mostly text-oriented and ignores the interpretation function of multimodal resources to educational knowledge and its rich characteristic representation. In order to better promote the follow-up research work, this paper reviews the educational knowledge graph from the perspective of multi-modal resources. Firstly, it introduces the concept and classification of knowledge graph. Secondly, it introduces the connotation of the educational knowledge graph, summarizes the definition, classification and construction framework of educational knowledge graph; Thirdly, combined with deep learning represented by neural network, it introduces the construction technology of educational knowledge graph. Finally, it summarizes the specific application of the educational knowledge graph, points out the problems existing in current research and future research directions. [ABSTRACT FROM AUTHOR]

Based on the typhoon optimal path data set from China Meteorological Administration during 2005 and 2020, and the Himawari-8 and FY4 satellites cloud image data, combined with the new technology of artificial intelligence, this paper applies deep learning models to typhoon vortex identification, typhoon location and intensity determination, typhoon intensity change detection, etc. To be specific, the new technology has been applied to develop a typhoon vortex identification model based on depth image target detection, a typhoon intensity intelligent determination model based on image classification and retrieval, and a typhoon rapid enhancement discrimination model considering the characteristics of time and space series. Tests with all the samples of vortices occurred in 2020 show that the correct identification rate of typhoon vortex with intensity reaching strong tropical storm or stronger is over 90%, the mean absolute error and root mean squared error of typhoon intensity estimation are 3.8 m/s and 5.05 m / s, respectively, and the comprehensive accuracy rate of annual independent sample intensity rapid enhancement estimation is 65.3%. The system realizes the automatic identification, positioning, intensity determination, and real-time intelligent tracking of tropical cyclones by using satellite images with high spatial and temporal resolution. [ABSTRACT FROM AUTHOR]

[Purpose/Significance] At present, most of the training models used in the research of news classification are non-active learning. There are common problems about these models, including data cannot be labeled immediately and the labeling cost is too high, which also hinders the analysis of agricultural news. Especially because of the explosive growth of news data in the network era, it is more difficult to label data, train supervised text classification models, and screen relevant news in the field of agriculture from diversified online news sources. In order to solve this problem, the most commonly used pool based active learning or deep active learning technique is used to select more valuable and representative data from unlabeled data for manual labeling, and construct labeled data sets to improve the efficiency and effect of news classification and agricultural news mining. [Method/Process] The commonly used machine learning models for text classification, such as random forest classifier, polynomial naive Bayes classifier and logistic regression classifier, were combined with the active learning method with the lowest confidence to analyze the effect, and the BERT model was combined with the three sampling strategies of discriminative active learning, deep Bayes active learning and lowest confidence for deep active learning training. On the news corpus of 19 847 samples crawled and cleaned by crawler technology from Sina and other news websites, aiming at screening agricultural related news from diversified news samples of various topics, the iterative experiment of adding 30 samples per round was tested to check the improvement effect of F1 score under various method combinations with the increase of the number of annotation. In addition, the representativeness and diversity of the samples selected by the sampling function of each method in the deep active learning method of the BERT model were compared, so as to understand the characteristics of each strategy and provide inspiration for the selection and improvement of Al strategy in the future. In addition, this paper also analyzed how much labeling cost can be saved by using the proposed method. [Results/Conclusions] When comparing a variety of machine learning models, it is found that although the gradient boosting tree and support vector machine classifier have high accuracy, they are not suitable for active learning because of their low efficiency in text data processing of large-scale high-dimensional data. After combining other machine learning models and the BERT model and training text models with the corresponding active learning or deep active learning methods, it is found that the application of active learning method can significantly improve the training process of each model. Among them, the BERT model, combined with discriminative active learning sampling function, has the best news text classification effect and the lowest annotation data requirements. The representativeness and diversity of the samples selected by discriminative active learning sampling function are also the highest, which explains the source of the advantages of this method. It can also be found that for the same task model, the higher the accuracy of classification is required, and the active learning method can save more annotation cost than non-active learning. [ABSTRACT FROM AUTHOR]

Objectives: Mars is the main target object for deep space exploration. Mars rovers, or roving probes, are important tools for surface exploration and scientific research on Mars. For the growing amount of remote sensing data collected by Mars rovers, there is an urgent need for a method that can intelligently detect novel targets of scientific value from the massive amount of images, reduce the time cost of detection planning, and provide information for subsequent scientific analysis. The traditional novel detection methods mostly include distance-based metrics and image-based reconstruction methods, distance-based metrics calculate novel scores pixel by pixel without considering spatial contextual information, and image-based re‐ construction methods focus on reconstruction of typical landscape backgrounds, and novelty is manifested by image reconstruction errors, which is not effective in extracting small novel targets such as boreholes and dust removal points. Methods: To address the above problems of traditional novel detection methods in Mars rover novel target detection, this paper proposes an improved Mars rover multispectral image depth novel target detection method, uses full convolutional self-coding neural network to extract deep features for typical landscape reconstruction, and joints Mahalanobis distance for novel target and typical landscape background separation, fully exploits the spatial and spectral dimensional features to improve the accuracy of Mars rover novel target detection results. Results: The experiments use the multispectral image dataset of Curiosity rover released by NASA ( National Aeronautics and Space Administration ), and the proposed convolution auto-encoder combined Mahalanobis distance method(CAE-M) is compared with Reed-Xiaoli detector, principal component analysis, convolution auto-encoder convolution, and generative adversarial networks on the surface of Gale crater. The results show that CAE-M outperforms previous detection methods in terms of detection accuracy and visual interpretation, and has a balanced and stable performance in different classes of novel target detection. Conclusion: The proposed CAE-M method comprehensively utilizes spatial and spectral information of multispectral images, which can help Mars rover exploration missions to quickly and intelligently screen and sort novel data with scientific value in massive data, save the time and cost of route planning, improve scientific returns. [ABSTRACT FROM AUTHOR]

Objectives: The scale of buildings and their distribution is key indicators to measure the economic and social development of a region. Therefore, it is significant to study the extraction of buildings based on remote sensing images. Existing neural network methods still have shortcomings in the completeness of building extraction and the accuracy of building edges. To solve the above problems, this paper proposes a multi-level feature fusion network (MFFNet) based on high-resolution images. Methods: Firstly, we use edge detection operators to improve the ability of the network to recognize the boundaries of buildings. Secondly, we use a multi-path convolution fusion module to extract building features from multiple dimensions, and introduce a large receptive field convolution module to break through feature extraction. The process is limited by the size of the receptive field. After fusing the extracted features, the convolutional attention module is used to compress them, and the global features are further mined by pyramid pooling, so as to achieve high-precision extraction of buildings. Results: The current mainstream UNet, pyramid scene parsing network (PSPNet),multi attending path neural network (MAPNet) and multiscale-feature fusion deep neural networks with dilated convolution (MDNNet) are used as the comparison methods, and we use Wuhan University Aerial Image Dataset, Satellite Dataset II (East Asia) and Inria Aerial Image Dataset as experimental data for testing. Compared with the other four methods, MFFNet improves intersection over union, precision, recall, F1-score and mean average precision by 1.53%, 2.65%, 2.41%, 3.32% and 1.19% on average, achieves a better effect. Conclusions: MFFNet not only accurately captures the detail features of buildings, but also strengthens the extraction and utilization of global features. It has better ex‐ traction effect on large buildings and buildings in complex environment. [ABSTRACT FROM AUTHOR]

China is the major producer and consumer of pork all around the world. In China, pork production has accounted for more than 60% of meat production for a long time. Nowadays, the pig industry has gradually shifted from decentralized and extensive breeding to intensive large-scale breeding, promoting production efficiency. At present, as people's demand for the quantity and quality of pork is increasing, China's pig industry faces the problem that the output and quality of pork can not meet people's daily needs. With the rise of artificial intelligence technology in recent years, deep learning technology has been developed rapidly and widely used in image and audio recognition, natural language processing, robotics, bioinformatics, chemistry, finance, and other fields. It is also an essential tool in developing precision livestock farming. The body condition, behavior and health status of pigs could directly affect the income level of pig farms, so through using deep learning technology, we can quickly and accurately acquire the relevant information about pigs, and carry out precise management to improve the feeding efficiency and welfare levels of pigs. This paper expounds on the research progress and application status of deep learning technology used in target pig detection, pig image segmentation, pig body condition and abnormal monitoring, as well as pig behavior recognition. Then we put forward the improvement strategy of deep learning technology used in the process of pig feeding, which make it easier for researchers to understand. At the same time, we summarize and analyze the data sources and datasets, application scope, and model optimization of previous works using deep learning technology in pig breeding. In the field of data sources and datasets, mobile computer vision systems are currently more suitable than systems with many fixed cameras when they are applied in pig houses, therefore, further research could focus on how to use deep learning technology to mobile computer vision systems. Deep learning technology requires a large amount of data to learn data features, which is a massive disadvantage for pig applications. To get pre-training weights suitable for agricultural data sets, a large number of public datasets and a unified dataset standard suited to the pig field should be established. In terms of application scope, due to the short application time of deep learning in the process of raising pigs, many critical occasions are not involved or seldom involved in existing research, so the application scope of deep learning should further expand. In model optimization, optimizing the model to meet the needs of practical application scenarios is the direction of future research. Optimizing the model to achieve a better balance between size and model performance requires further study. Optimizing the model to locate the start and end time of the behavior in unclipped video and recognize the behavior of target pigs is the challenge in identifying the behavior of pigs in future. China's research on deep learning is at the top level, so the application prospect of deep learning in pig farming is highly expected. Combined with the actual production scene of pigs, deep learning technology will significantly contribute to the development of precision livestock farming. [ABSTRACT FROM AUTHOR]

Salient object detection (SOD ) aims at detecting the significant parts in the images automatically and efficiently, which plays an important role in computer vision tasks such as object detection and recognition, key points location, visual tracking, semantic segmentation and so on. This paper focused on a comprehensive review of salient object detection, which covered the closely related fields, models, datasets, applications and future directions. It introduced the relationship between salient object detection and closely fields at first. Then, it analyzed salient object detection algorithms in terms of flowcharts, ideas, innovations, performance as well as applicability. Thirdly, it introduced the public datasets which improving continuously with the development of the saliency detection models. Finally, it discussed a prospect of future researches and several real application scenes of salient object detection. [ABSTRACT FROM AUTHOR]

in order to get better performance of 3 D model retrieval and classification,this paper proposed the characterization of 3 D models by multi-modal information fusion,which based on deep-learning model.The method considered the correlation of different modalities in the training step for extracting a more robust feature vector,which benefitted from the proposed correlation loss function.In addition,it applied fusion features to 3 D model retrieval and classification.This paper evaluated the proposed method on the ModelNet40 dataset for 3 D models classification task and retrieval.The comparison between the experimental results and the existing methods proves the superiority of this method,which provides a new idea for the field of 3 D model retrieval and classification. [ABSTRACT FROM AUTHOR]

Aiming at the problems of low accuracy and poor generalization of industrial aluminum surface defect detection, this paper proposes a multi task feature fusion detection method based on improved segmentation decision network. Firstly, aiming at the common pit defects in industry, it constructs the defect detection data set of this paper; secondly, on the basis of the original segmentation detection model,it also adds the aluminum area segmentation network to reduce the demand for samples in the network structure; in addition, the Squeeze-and-Excitation module is introduced to make the model focus more on the dimensional features related to aluminum defects and improve the detection accuracy of the network. Finally, a comparative experiment is carried out on the data set of dirty spots on the aluminum surface. The structure shows that the segmentation accuracy of the proposed method is improved by 1.6% and the classification accuracy is improved by 4.01% compared with the original model; compared with FCN and u-net network, the detection accuracy of image segmentation is 0. 76% higher than that of FCN and U-Net network, and 1. 12% higher than that of VGG and Resnet50 network. The experimental results have proved that this method has sufficient detection accuracy in the detection of dirty spots on the aluminum surface, and is stable and efficient. [ABSTRACT FROM AUTHOR]

Traditional image inpainting algorithms cannot accurately capture high-level semantics when regions involve complex non-repetitive structures, such as faces. In the past three years, the method based on deep learning has been applied to image inpainting, and the structural similarity of the repair results had increased by more than 10% compared with the traditional methods. This paper firstly expounded the research and development process of facial completion technologies, mainly introduced the face repair algorithm based on deep learning, which was divided into unsupervised and supervised categories. In each category, it focused on the various facial completion algorithm ideas that had emerged in recent years. Then it summarized the six mainstream types of image datasets. Two evaluation indexes were summarized in this paper. Finally, it discussed the future research direction of facial completion technique. [ABSTRACT FROM AUTHOR]

In distributed video coding ( DVC ), generating high-quality side information and improving decoding performance in various sports scenes is a crucial research area. This paper proposed a deep learning optical flow interpolation ( OFI ) side information generation algorithm. The encoder used interval overlapped distributed arithmetic codes( DAC) to encode the video. The decoder extracted the decoded key frames and inputted them into the optical flow interpolation network for side information generation. The network used a multi-layer optical flow module to generate optical flow, and used the optical flow to warp key frames to generate a preliminary side information estimation, and then underwent a fusion process to eliminate occlusion to produce more refined results. Finally, the side information assisted the decoding tree to complete the decoding. The experimental results show that comparing with the existing method, the PSNR of this method increases up to 2. 25 dB, which happens mainly in the linear motion scenes. At the same time, the SSIM increases 0. 001 5 - 0. 064 8 in both linear and non-linear scenes. The decoded video rate-distortion curve shows consistent results. It proves that the algorithm can improve the estimation of linear motion side information, and it also has excellent robustness on the structure of non-linear motion side information. [ABSTRACT FROM AUTHOR]

Due to the low randomness of character distribution and high randomness of word combination in word-based DGA domain names, the traditional detection methods are difficult to identify them. The method based on deep learning such as LSTM can capture domain's sequence features, but it lacks the word combination features and has low detection accuracy. In order to solve above problems, this paper proposed a new DGA domain name detection method based on fusion embedding layer. In domain embedding stage, it carried out a fusion vector representation based on characters and word segmentation technology . Then, it constructed a hybrid deep learning model based on CNN and BiGRU. The experimental results show that the proposed model has increased by 3 . 1 % and 4. 3% respectively in accuracy compared with the models based on CNN or LSTM alone. When dealing with the matsnu, suppobox and ngioweb DGAs, the proposed model is more effective. [ABSTRACT FROM AUTHOR]

Generating images from text is a challenge task in machine learning community. Although significant success has been achieved so far, problems such as unstable network training and disappearing gradients still exist. In response to the above shortcomings, based on the stacked generative confrontation network model (StackGAN), this paper proposes a text-to-image generation method that combines spectral normalization and perceptual loss function. Firstly, the network model applies spectral normalization to the discriminator, restricts the gradient of each layer of the network to a fixed range, slows down the convergence speed of the discriminator, and hence improves the stability of network training. Secondly, the perceptual loss function is added to the generator network to enhance the consistency between the text content and the generated image. The network model uses Inception scores to evaluate the quality of the generated images. The experimental results show that, compared with the original StackGAN, the network model has better stability and generates clearer images. [ABSTRACT FROM AUTHOR]

Objectives: Manual seismic phase picking and earthquake catalog building become challenging because the number of seismic stations increases. With the rapid development of artificial intelligence, several deep learning models are developed for seismic phase picking, but their performance and applicability remain ambiguous. Methods: Phase-Net and EqT can extract relevant features at different levels from raw seismic waveforms and map them into probabilistic sequences. In this paper, the performance of these two widely-used deep learning models for earthquake detection and phase picking is investigated with the 2019 Ridgecrest earthquake (California, USA) sequence and the 2019 Weiyuan Ms 5.4 earthquake (Sichuan, China) sequence. Besides, the completeness and accuracy of the constructed earthquake catalogs are analyzed and evaluated by cross-validation. Results: The results show that the two models, though trained with datasets from different regions, perform well in cross-region applications and are comparable to manual analysis. EqT1 provides high detection precision with few false alarms, even under a low threshold. But EqT1 has a limited recall rate, implying that fewer earthquakes are detected compared with Phase-Net. The picking performance of Phase-Net is sensitive to the employed threshold. With the low and moderate thresholds, Phase-Net can detect more earthquakes than EqT1 and manual catalog, while registering a relatively low detection precision. With the high threshold, the precision of Phase-Net is comparable to that of EqT1. EqT2 is not recommended due to its inferior performance over EqT1 and Phase-Net. Conclusions: EqT1 provides a higher precision when being employed for seismic phase picking, leading to earthquake catalogs with better accuracy. Phase-Net has a higher recall rate and generates more complete catalogs. Both models can facilitate automatic and rapid construction of earthquake catalogs. [ABSTRACT FROM AUTHOR]

Based on the importance of helmet detection in construction site management and the cost control of hardware facilities in engineering projects, this paper proposes a helmet detection approach based on Lighter and Tiny-YOLO ( LT-YOLO), a lightweight and improved version of the deep learning network Tiny-YOLO v3. The number of prediction layer is increased and an innovative R-DSC feature extraction module is introduced in LT-YOLO. The complexity of the model can be greatly reduced by R-DSC module without changing the size of the network inputs and outputs. The experimental results showed that LT-YOLO achieved an excellent balance between light weight and detection performance, reaching 59.3 mAP (mean average precision) and 59.4% Recall with only 3.5 M parameters. Because of very few parameters and very low computation, LT-YOLO has low dependence on high computing hardware, and is suitable for actual construction site safety management. LT-YOLO can greatly reduce the cost of enterprises and improve the level of construction safety management. [ABSTRACT FROM AUTHOR]

With the rapid development of deep learning and 3 D sensing technology, point cloud classification has been widely used in intelligent classification and other fields. In order to better promote the research and application of point cloud classification technology, this paper combed, analyzed and summarized the research progress of related methods systematically by using pipeline architecture. Firstly, according to the different point cloud data processing methods, it analyzed and summarized the existing point cloud classification methods into indirect point cloud based methods and direct point cloud based methods. Then, it introduced the representative methods and the latest research, compared and analyzed the core ideas, advantages and disadvantages, scope of application, application scenarios and experimental results of the main methods. Finally, it prospected the future development and research direction of point cloud classification from four aspects. The results show that 2D-3D feature fusion of indirect and direct point cloud methods is an important development direction in the future. [ABSTRACT FROM AUTHOR]

CD56 is a nerve cell adhesion molecule that can be used in the diagnosis and study on a variety of tumor cells. CD56 is one of the latest tumor molecular markers, and the research on CD56 images in the field of digital medical image processing has just started. With the development of deep learning technologies such as semantic segmentation, more and more researchers are applying semantic segmentation technology to medical image processing in order to assist doctors in diagnosis. In CD56 images, the proportion of the number of pixel points of background, negative cells and positive cells is very unbalanced, which is roughly 70∶10∶1 and affects the segmentation effect of semantic segmentation technology on CD56 images. In this paper, the loss function of the semantic segmentation model is improved by adding loss weight to different class and adding adaptive weight to each pixel, so that the model can pay more attention to cells, especially positive cells. At the same time, the clustering method is used to refine the annotations of CD56 images before model training, which further improves the segmentation accuracy of the model. The experimental results on CD56 image dataset show that, the refinement of the image annotations and the improvement of the loss functions in the relevant semantic segmentation models can improve the segmentation accuracy of the CD56 images. [ABSTRACT FROM AUTHOR]

When traditional convolutional neural networks are used in mineral species identification, due to their large parameters and the limitations of fixed input image resolution, sufficient computing resources and image preprocessing are required, which are difficult to deploy in actual exploration. For this reason, based on the depth separable convolution,combined with the attention mechanism, this paper constructs a mineral intelligent recognition model through dense connection, and this model can train multi-resolution mineral images. Ex perimental results show that the model's memory usage is only 20 Mb, the verification accuracy and the test accuracy are both higher than 90 %, the classification effect is better than the classic convolutional neural network, and it shows excellent ability to distinguish positive and negative samples. These results show that the proposed model h as fewer parameters and low memory footprint which is suitable for portable devices. This model can successfully identify mineral images with different resolutions, which has good generalization and potential application. [ABSTRACT FROM AUTHOR]

In order to solve the problem of mismatch in cross-domain medical image analysis, this paper proposed an UAL-DAF. Specifically, the framework used the ATM and the STM that combined conditional generative adversarial networks to reduce the differences in the appearance and semantic levels of cross-domain medical images, respectively. Finally, in the challenging medical image segmentation, the results are significantly better than existing methods. Therefore, the framework can extract the appearance and semantic level information of domain adaptation knowledge, and realize the collaborative integration of domain knowledge. [ABSTRACT FROM AUTHOR]

Non-small cell lung cancer (NSCLC) accounts for 80%-85% of lung cancer and is one of the malignant tumors that seriously endanger human health. Treatment of early NSCLC is based on surgical excision or stereotactic body radiation therapy. Whether there is regional lymphatic metastasis when diagnosis is confirmed will affect the choice of local treatment, and whether there is still a risk of lymphatic and hematologic metastasis after the completion of local therapy will be the basis for accurate decision of adjuvant therapy. How to predict the risk of lymphatic or hematologic metastasis of NSCLC remains a challenge. With the development of tumor and the plasticity of treatment, heterogeneity of biological characteristics of tumors in time and space seriously affects the accuracy of clinical diagnosis, treatment and prognostic prediction. Due to the heterogeneity of tumor, it is difficult for invasive biopsy to show the full picture of tumor biological characteristics as a gold standard, which promotes clinical attention to non-invasive methods, such as medical images, to identify biological features. The method to identify tumor biological features based on medical images experiences from the qualitative analysis of artificial visuals to the modeling of advanced statistical methods for manual extraction of imaging features, and then to the application of radiomics and deep learning models, which provide new possibilities for accurate and efficient medical imaging analysis. Based on chest computed tomography (CT) imaging, this paper summarized the progress of research on the prediction of risk of lymphatic and hematologic metastasis, an important factor affecting early NSCLC treatment decision-making. [ABSTRACT FROM AUTHOR]

Objectives: CAE (convolutional autoencoder) combines the autoencoder neural network structure and the convolutional network structure. This paper processes the gravity data based on the deep learning, and reconstructs the gravity contour image from the noisy gravity data with the CAE. The autoencoder structure contains the equal dimensional input and output data, which goes for the gravity data processing. The convolutional network structure is widely used in image recognition recent years, which can learn and recognize the specific objects in an image. Methods: To create the training set, we generate 1000 hexahedrons with the random triaxial length parameters and the random density to simulate the natural gravity source bodies. The 2D gravity data set for these bodies is fast computed with the gravity forward formula and the noisy gravity data set is generated by adding Gaussian noise to the computed data. Meanwhile the accuracy of the length parameters and density is limited in order to improve the representativeness of the training data set. Hence the 2D gravity data and the noisy data will be output and input data of the CAE. We design 5 layers CAE. The input and output layers both are 26×26. 32 and 64 feature maps are generated with 3×3 convolution kernels in the inner layers. The training is executed using RMSProp(root mean square propagation) optimizer. Results: To test the generalization of the trained CAE, the testing set with 500 samples is generated in the same way with the training set. The relative error histogram on the testing set shows that the reconstructing error is less than 5% and most of the error is around 0. To test the recognition of the gravity features in the 2D gravity, we test the CAE with simulated gravity data which contain 2 and more gravity anomalies. The results show that the CAE can recognize all gravity anomalies in an noisy gravity image and reconstruct them and the output image shows fairly fine reconstruction from the gravity with 10% noise. The measured gravity data of Kauring testing ground are used to test the generalization performance of the CAE. The reconstructed image is compared with the traditional FFT(fast Fourier transform) filtering and db wavelet filtering results. To increase the reconstruction difficulty, the Kauring data is added with 10% Gaussian noise. The results show that the trained CAE performs significantly less error than the FFT filter and the wavelet filter. Conclusions: In terms of reconstructing performance, the CAE can avoid over processing that may eliminate the real gravity anomaly, which is often seen in filtering especially for FFT filtering. The CAE reconstructs all of the 3 gravity anomalies of Kauring while the FFT filtering reconstructs 2. And the over-smoothing artifact is less than filtering methods. The CAE can process 2D gravity data and show lower error and generate less artifacts than filtering methods in our tests with testing set, simulated gravity data and measuring data. Beyond that, this article reveals that the deep learning can learn gravity image. With the convolution structure in a neural network, the network can learn single gravity anomaly samples and extensively recognize all gravity anomalies in a gravity image, which signifies that the gravity can integrate with the artificial intelligence technology. [ABSTRACT FROM AUTHOR]

At present, super-resolution method based on CNN can obtain excellent results in the evaluation metric of peak signal- to-noise ratio and the structural similarity index, but the visual perceptual quality of super- resolution image is poor, and the details of the facial features are lost. In order to solve these problem, this paper designed a new deep neural network to predict the super-resolution wavelet coefficients to get clear super-resolution face images . Firstly, it used the prior knowledge of face images to manually give the facial features more attention. Then it introduced linear low-rank convolution in the network, and finally used the idea of long-distance dependence to supplement the details of the super-resolution images. The experimental results show that the method in this paper has achieved competitive results in the evaluation metric of peak signal-to-noise ratio and structural similarity index, and the visual perceptual quality of its super-resolution face images are excellent. [ABSTRACT FROM AUTHOR]

To solve the problems that the objects with different semantic classes are identified as co-salient objects in current co-saliency detection methods, this paper proposed a CNN and semantic correction-based co-saliency detection method (CSCCD). The proposed method first adopted the guided super pixel filter to process the super pixels obtained by SLIC and the saliency results generated by DSS, which showed clear object boundaries. Then it utilized Mask R-CNN to extract semantic features. It proposed the definitions of image semantic feature and semantic consistency. It also defined the image group semantic correction to solve the problem that detected the objects with different pose belonging to a semantic class as different semantic classes. With the concept, this paper defined image group semantic correlation class, solving semantic correlation problem of multiple images. It generated the final co-saliency detection results by fusing the saliency detection regions with the image group semantic consistent regions. The experimental results on public benchmark datasets show that this algorithm can effectively highlight the whole and outline of the object, and its comprehensive performance in objective quantification is obviously improved. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the existing super-resolution algorithm based on deep learning didn't make full use of the feature information of each level, resulting in low reconstruction accuracy and large parameter quantity, this paper proposed a double dense connection structure named densely channel attention skip connection network. In the inner structure of the network, it improved the original dense cascade block to generate a channel separable dense cascade block. The outer structure adopted a densely residual connection and attention mechanism to fuse the features extracted by the dense block to achieve the goal that less convolution layer and higher precision effect. This paper tested the network models on several benchmark datasets . The results show the proposed model has higher accuracy and fewer parameters than the other models. [ABSTRACT FROM AUTHOR]

Aiming at the problem of less research and low detection rate of malware detection in the current container environment, this paper proposed a static detection method of malware in container based on LS TM-CNN. The purpose of the method is to detect malware before it runs, to block its attack behavior from the source, and to reduce the performance loss brought by the detection process to the operation of container. The method obtained the software to be tested in the container without agent, and extracted API( application programming interface) call sequences as program behavior data, used word2vec model to vectorize API call sequences, and extracted the semantic information and multi-dimensional local features based on LSTM and CNN respectively to detect malware. This paper implemented the method in container environment and tested on an open sourced dataset VirusShare, the results show that the method can achieve a detection rate of 99. 76% and the false alarm rate is less than 1%, which is superior to other similar methods. [ABSTRACT FROM AUTHOR]

Aiming at the problem of poor diagnosis ability and generalization ability of the trained model caused by serious imbalance of bearing fault data, this paper proposed a method of generative adversarial networks based on Wasserstein distance to balance dataset. Firstly, it trained a small number of fault samples for adversarial training. Then when the network reached the Nash equilibrium, it added the generated fault samples to the original small number of fault samples to balance the dataset. This paper proposed a diagnostic model based on global average pooled convolutional neural network. The balanced data set was input into the diagnostic model for training. The model was adaptively extracted layer by layer to achieve accurate classification diagnosis of faults. The experimental results show that the proposed diagnostic method is superior to other algorithms and models, and has strong generalization ability and robustness. [ABSTRACT FROM AUTHOR]

At present, multi-agent reinforcement learning algorithms mostly adopt frameworks that are centralized in learning and decentralized in action. These frameworks may take too long to converge or may not converge at all. In order to speed up the collective learning time of multi-agents, this paper proposed a novel multi-agent group learning strategy. It used recurrent neural network (RNN) to predict the grouping matrix of multi-agents to share the experience between them, resulting in improved learning efficiency within the multi-agents group. Meanwhile, this paper proposed the concept of information trace to remedy the problem that the agents could not share information brought by the grouping. In order to strengthen the retention of excellent experience within the group, this paper proposed the practice of delaying the death time of excellent agents in the group. Finally, the results show that, compared to MADDPG, the training time is reduced by 12% in the labyrinth experiment and by 17 % in capture the flag experiment. [ABSTRACT FROM AUTHOR]

In order to understand the development of outlier detection technique detailly, this paper overviewed the outlier detection techniques in recent years. Firstly, it introduced and summarized the definition of outliers, the causes of outlier and classification of the outlier mining algorithms. Next, it overviewed outlier detection methods based on neighbor, based on distributed framework and based on deep learning. Especially, it discussed the most representative methods in outlier detection, and pointed out the merits and drawbacks. In the end, this paper concluded and discussed the future direction of outlier detection techniques. [ABSTRACT FROM AUTHOR]

Based on video action recognition, this paper firstly reviewed the traditional action recognition algorithms based on RGB, including traditional algorithms and deep learning algorithms. Because of the influence of background lighting, RGB-videobased action recognition is low accuracy, but it has rich color appearance information. Then this paper analyzed and summarized action recognition algorithms based on RGB-D, which was divided into depth sequence, bone and multi-feature fusion. RGB-D video has multiple modes and can provide more information for action recognition, which performs better than RGB video but also brings new challenges. Finally, this paper prospected the common data sets and the possible future development direction. [ABSTRACT FROM AUTHOR]

With the advancement of data acquisition technology, spatiotemporal data with geographic location information are growing rapidly, so it is urgent to explore effective spatiotemporal data modeling methods. Spatiotemporal sequence prediction is one of the basic methods in spatiotemporal data modeling and widely used in many fields. At present, there is no Chinese literature on its review, so it is of great significance to summarize these methods. For the spatiotemporal sequence prediction problem, this paper reviewed its application background and development history. Then it introduced some related definitions and characteristics. According to its category, the spatiotemporal sequence prediction methods, based on traditional methods, classical machine learning methods and deep learning methods are analyzed respectively as well as their application scope, advantages and disadvantages. Finally, this paper made a review and prospect over the research directions of spatiotemporal sequence prediction, laying a theoretical foundation for researchers to further study this problem [ABSTRACT FROM AUTHOR]

The automatic detection of mesoscale ocean eddies is extremely essential to monitor their dynamic changes. Therefore, effective detection of ocean eddies is vital to improve understanding of ocean dynamics. The traditional methods of detecting eddies using remote sensing data are usually based on physical parameters, geometric features, manual features or expert knowledge. In recent years, the deep learning technology has been improved by many experts. In this paper, the deep learning method is used to detect the ocean eddies from the sea surface height (SSH) data. Firstly, a multi-eddy detection model based on deep convolution neural network is proposed aiming to resolve eddy detection challenge on SSH data photoed by satellite. The model can accurately extract the feature information of eddies and fit the relationship between semantic information and sea level anomaly. Secondly, a new dataset, i.e., SCSE-Eddy, is used to train the proposed model and evaluate the performance of eddy detection method based on artificial intelligence (AI). This dataset is composed of the daily satellite remote sensing SSH data covering the South China Sea and its eastern sea area over the past 15 years. Experimental results show that, compared with the existing methods, the model proposed in this paper achieves the best performance and distinguishes close eddies well. [ABSTRACT FROM AUTHOR]

Accurate detection and recognition of plant diseases is the key technology to early diagnosis and intelligent monitoring of plant diseases, and is the core of accurate control and information management of plant diseases and insect pests. Deep learning can overcome the disadvantages of traditional diagnosis methods and greatly improve the accuracy of diseases detection and recognition, and has attracted a lot of attention of researchers. This paper collected the main public plant diseases image data sets all over the world, and briefly introduced the basic information of each data set and their websites, which is convenient to download and use. And then, the application of deep learning in plant disease detection and recognition in recent years was systematically reviewed. Plant disease target detection is the premise of accurate classification and recognition of plant disease and evaluation of disease hazard level. It is also the key to accurately locate plant disease area and guide spray device of plant protection equipment to spray drug on target. Plant disease recognition refers to the processing, analysis and understanding of disease images to identify different kinds of disease objects, which is the main basis for the timely and effective prevention and control of plant diseases. The research progress in early disease detection and recognition algorithm was expounded based on depth of learning research, as well as the advantages and existing problems of various algorithms were described. It can be seen from this review that the detection and recognition algorithm based on deep learning is superior to the traditional detection and recognition algorithm in all aspects. Based on the investigation of research results, it was pointed out that the illumination, sheltering, complex background, different disorders with similar symptoms, different changes of disease symptoms in different periods, and overlapping coexistence of multiple diseases were the main challenges for the detection and recognition of plant diseases. At the same time, the establishment of a large-scale and more complex data set that meets the specific research needs is also a difficulty that need to face together. And at further, we point out that the combination of the better performance of the neural network, large-scale data set and agriculture theoretical basis is a major trend of the development of the future. It is also pointed out that multimodal data can be used to identify early plant diseases, which is also one of the future development direction. [ABSTRACT FROM AUTHOR]

Diseases and pests are main stresses to crop production. It is necessary to accurately and quickly monitor and control the stresses dynamically, so as to ensure the food security and the quality and safety of agricultural products, protect the ecological environment, and promote the sustainable development of agriculture. In recent years, with the rapid development of the unmanned aerial vehicle (UAV) industry, UAV agricultural remote sensing has played an important role in the application of crop diseases and pests monitoring due to its high image spatial resolution, strong data acquisition timeliness and low cost. The relevant background of UAV remote sensing monitoring of crop disease and pest stress was introduced, then the current methods commonly used in remote sensing monitoring of crop disease and pest stress by UAV was summarized. The data acquisition method and data processing method of UAV remote sensing monitoring of crop disease and pest stress were mainly discussed. Then, from the six aspects of visible light imaging remote sensing, multispectral imaging remote sensing, hyperspectral imaging remote sensing, thermal infrared imaging remote sensing, LiDAR imaging remote sensing and multiple remote sensing fusion and comparison, the research progress of remote sensing monitoring of crop diseases and pests by UAV worldwide was reviewed. Finally, the unresolved key technical problems and future development directions in the research and application of UAV remote sensing monitoring of crop disease and pest stress were proposed. Such as, the performance of the UAV flight platform needs to be optimized and upgraded, as well as the development of low-cost, lightweight, modular, and more adaptable airborne sensors. Convenient and automated remote sensing monitoring tasks need to be designed and implemented, and more remote sensing monitoring information can be obtained. Data processing algorithms or software should be designed and developed with greater applicability and wider applicability, and data processing time should be shortened by using 5G-based communication networks and edge computing devices. The applicability of the algorithm or model for UAV remote sensing monitoring of crop disease and pest stress needs to be stronger, so as to build a corresponding method library. We hope that this paper can help Chinese UAV remote sensing monitoring of crop diseases and pests to achieve more standardization, informatization, precision and intelligence. [ABSTRACT FROM AUTHOR]

CNN (convolutional neural network) is widely used in image classification tasks. Most existing CNN-based classification models are trained as flat N-way classifiers. However, the difference among different image categories limits the capacity of the classifier. To solve the problem, this paper proposed CT-CNN (confusion tree convolutional neural network) model which combined the CT with CNN. CT-CNN first established a CT to identify the confused categories. Then it embedded the hierarchical structure of CT into CNN model, which leaded the CNN training procedure to pay more attention on strongly confused categories. Experiments on public datasets prove that CT-CNN can overcome the limitation of uneven distribution of classification difficulty between categories of large-scale datasets and achieve better performance on complex large-scale image classification tasks. [ABSTRACT FROM AUTHOR]

Existing deep learning based single image super-resolution (SISR) models usually improve the fitting ability of the model by increasing the number of network layers, but fail to fully extract and reuse features, leading low quality of reconstructed images. To solve this problem, this paper proposed an image super-resolution model based on feature fusion and attention mechanism. This model used residual in residual (RIR) structure in feature extraction module. The feature extraction module of the network consisted of several residual groups. Each residual group consisted of several residual blocks. This module implemented local feature fusion in each residual group and global feature fusion between each group. In addition, this model introduced coordinate attention module into each residual block and spatial attention module into each residual group. It verifies that the model is able to fully extract features and reuse features. The final experimental results show that the model is superior to the existing models in objective evaluation indexes and subjective visual effect. [ABSTRACT FROM AUTHOR]

This paper proposed an optimization model of deep-knowledge tracking（KTSA-BiGRU） based on self-attention mechanism and bidirectional GRU neural networks owing to the existing deep-knowledge tracking models with weak capture of complex relationships between input exercises and inability to effectively handle long-sequence input data. Firstly, it mapped the learner’s historical learning interaction sequence data to the real value vector sequence. Then, it trained the bidirectional GRU neural network as input to model the learner’s learning process, and finally used, the self-attention mechanism to calculate the probability of the learner correctly answering the next question based on the hidden vectors of the bidirectional GRU neural network output and the attention weight. The performance analysis on the three public datasets can improve the prediction accuracy of deep knowledge tracking. [ABSTRACT FROM AUTHOR]

Aiming at the problem of low person re-identification accuracy caused by similar appearance and posture of person in surveillance videos, this paper proposed a video-based person re-identification method based on a graph model, which effectively utilized the time sequence information in the video to realize the information interaction across frames and intra-frame regions. Specifically, it used the correlation information between the cross-frame block regions to establish the inherent relationship between the regional nodes, and iteratively updated the regional information through feature propagation. On the other hand, in the metric learning process, it proposed a weighted loss function strategy, which optimized the binary allocation method (that is, discarding or retaining the sample) in the previous mining strategy into a continuous score allocation method, which solved the problem that the available samples were not used efficiently. Finally, it evaluated the model on MARS and DukeMTMC-VideoReID datasets, and the experimental results confirm the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

In order to sort out the development of deep learning methods in the field of human action recognition, this paper summarized the most representative models and algorithms in this field in recent years. Firstly, it described in detail the latest achievements, advantages and disadvantages and network structure of deep learning methods in video pre-processing stage based on the task flow of human action recognition. Then, it introduced two kinds of datasets related to human action recognition. Finally, it discussed and prospected the future research direction of human action recognition. [ABSTRACT FROM AUTHOR]

Occlusion in a real environment is one of the main obstacles to accurately analyze and recognize facial expressions. In recent years, researchers have used deep learning technology to solve the problem of high misrecognition rate of facial ex-pressions under occlusion conditions. It mainly summarized the deep learning algorithm of occlusion facial expression recognition and occlusion-related issues. Firstly, this paper summarized the development status of facial expression recognition under partial occlusion conditions, expression representation, and the data set used to study occlusion expression. Secondly, it re-viewed the occlusion expression identify the latest developments in deep learning methods and analyzed the impact of occlusion on expressions. Finally, it summarized the main technical challenges, research difficulties and possible coping strategies. The purpose is to provide a more useful reference and benchmark for future research on occlusion expression recognition. [ABSTRACT FROM AUTHOR]

Dialogue system aims to realize the communication for information between machine and human so as to provide personalized service, it has certain research value. In order to follow up the research in related fields timely, this paper reviewed the current research progress. Firstly, it introduced the development and classification of dialogue system. Secondly, it summarized the architecture of task-oriented dialogue system, focusing on module based and the end-to-end based frameworks, then summarized the advantages and disadvantages of the models. Thirdly, it introduced the evaluation methods and application fields of task-oriented dialogue system, including e-commerce, education, medicine and so on. Finally, it analyzed and summarized the problems and challenges faced by the multi-turn task-oriented dialogue system. [ABSTRACT FROM AUTHOR]