Search

Deep support vector data description (DeSVDD) is an emerging anomaly detection method based on the deep learning methodology. However, few studies take the confidence of DeSVDD predictions into account so that the present DeSVDD models cannot indicate the reliability degree of the anomaly detection results. In this paper, we enrich the theory of DeSVDD by building the model confidence definition and developing the corresponding calibration strategy. For one thing, by revisiting the methodology of DeSVDD-based anomaly detection, the confidence of detection results is presented to indicate if the prediction of DeSVDD is reliable. For another, we propose a modified power T -scaling strategy to smooth the anomaly scores of DeSVDD model and improve its calibration performance without changing the original detection results. Six open experiment datasets are used to perform the method testing and the experimental results confirm the effectiveness of our proposed calibration strategy. [ABSTRACT FROM AUTHOR]

Process mining focuses on the analysis of event log data to build various process analytical capabilities. Predictive process analytics has emerged as one of such key capabilities and it uses machine learning techniques to construct process prediction models. In recent years, deep neural networks have gained increasing interest in process prediction since they can handle multi-dimensional sequential inputs with minimal information loss. However, they are considered black-box models and existing studies in explaining deep neural network-based process predictions rely on only event-level features for explanation. In this paper, we propose a new approach for generating explanations for process outcome predictions at multiple levels. The approach is underpinned by three different prediction models: a transparent model for generating global explanations based on case-level features, an attention-based deep neural network for generating local explanations based on event-level features, and a novel eXplainable Dual-learning Deep network (XD 2 -net) for generating local explanations based on case-level features. Using three publicly available datasets, we have tested the applicability of the approach and further examined the multi-level explanations generated by the approach through an elaborate case study. Unlike others, the design of our approach promotes the idea of leveraging the complementary capabilities of different models and utilizing their strengths, rather than focusing on model performance competition. This will contribute towards generating more comprehensive explanations that meet the needs of different end users and purposes in the future. [ABSTRACT FROM AUTHOR]

With the growing complexity of onboard sensors and the widespread deployment of road sensors, deep learning enables fine-grained traffic prediction using massive amounts of raw traffic data, which facilitates accurate analysis of traffic information in the Internet of Vehicles (IoV). However, most existing studies focus on using all the local data to jointly build a prediction model, facing severe challenges of data security and privacy concerns as well as substantial communication overhead. To address these challenges, in this paper, we propose the Spatial-Temporal Traffic Prediction Network based on federated learning (F-STTP-Net), which only updates the model parameters to the centralized server without any private data. Firstly, we design a sub-area division method, which divides the road network into sub-areas with different macroscopic fundamental diagram properties. Then, we propose a local training model for each sub-area, which uses the graph attention network (GAT) and the long short-term memory (LSTM) to capture the spatio-temporal dependence of the road network. The model uses the branch structure to predict the traffic volume of each intersection in the sub-area. Finally, the local models are aggregated based on federated learning to form a powerful central model, which bridges the constraints on global data sharing and privacy guarantee. We conduct experiments on the real-life dataset in Xuchang Lotus Lake 5G automated vehicles demonstration area to demonstrate that F-STTP-Net can achieve excellent prediction performance without the interaction of sub-area raw data. In addition, the proposed model has a strong generalization ability and can be quickly transferred to a new sub-area. [ABSTRACT FROM AUTHOR]

Accurate multi-node load forecasting is the key to the safe, reliable, and economical operation of the power system. However, the dynamic nature of load and the coupling nature of networks are difficult to extract, making consistent and accurate forecasting of node load rather difficult. In this regard, this paper proposes a soft sharing multi-task deep learning method for multi-node load forecasting in the power system. It has the following aspects: (1) Considering the coupling characteristics of the node network, a multi-modal feature module, based on the inception strategy and gated temporal convolutional network (GTCN), is firstly designed to explore the coupling features implied in the node load data. (2) A novel multi-objective neural network model is proposed to achieve simultaneous prediction of multi-node load by integrating the multi-modal feature module and gated recurrent unit (GRU). For sharing the learning information of sub-networks, this paper uses the soft sharing mechanism to capture load features, which can better optimize the prediction task for each node load simultaneously. Load data from the New Zealand distribution network and AEMO are used to compare the proposed model's performance in various scenarios using regression metrics such as mean absolute percentage error (MAPE), Weighted Mean Accuracy (WMA), root mean squared logarithmic error (RMSLE), and Diebold–Mariano (DM). The simulation results show that the proposed method can explore the spatial–temporal coupling characteristics in multi-node load data. Compared with existing state-of-the-art multi-node load prediction methods, our proposed method's MAPE decrease 17.04% and 3.92% in Non-aggregation and Aggregation situations. • A novel deep multi-task learning method is proposed for the multi-node forecasting problem. • An IGTCN module is designed to extract the coupling features from different node loads. • The multi-task deep neural network exhibits the accuracy of multi-node load forecasting. [ABSTRACT FROM AUTHOR]

Since the generative adversarial network (GAN) was proposed in 2014, it has rapidly become a hot topic in the field of deep learning. In recent years, there are many optimizations for GAN, which are divided into three kinds, including the optimization of loss functions, external structure of network and internal structure of network. Few people optimizes GAN from internal structure of network, and because of the process of game, slow training speed is one of the biggest problems of GAN. This paper introduces the ideology of broad learning algorithm (Chen and Liu, 2017) to put forward the multi-judge generative adversarial network method based on different random features to enhance the quality of generative result and increase the training speed of GAN (Goodfellow et al., 2014). This model provides a method which reduce the input information of each discriminator to optimize the training process of GAN. The experiments on cifar10 and anime face dataset find that our model obtains a better performance than the GMAN model and Base model. This paper finds a group of hyper-parameters to enhance the BroadGAN. [ABSTRACT FROM AUTHOR]

Deep Reinforcement Learning systems are now a hot topic in Machine Learning for their effectiveness in many complex tasks, but their application in safety-critical domains (e.g., robot control or self-autonomous driving) remains dangerous without mechanism to detect and prevent risk situations. In Deep RL, such risk is mostly in the form of adversarial attacks, which introduce small perturbations to sensor inputs with the aim of changing the network-based decisions and thus cause catastrophic situations. In the light of these dangers, a promising line of research is that of providing these Deep RL algorithms with suitable defenses, especially when deploying in real environments. This paper suggests that this line of research could be greatly improved by the concepts from the existing research field of Safe Reinforcement Learning, which has been postulated as a family of RL algorithms capable of providing defenses against many forms of risks. However, the connections between Safe RL and the design of defenses against adversarial attacks in Deep RL remain largely unexplored. This paper seeks to explore precisely some of these connections. In particular, this paper proposes to reuse some of the concepts from existing Safe RL algorithms to create a novel and effective instance-based defense for the deployment stage of Deep RL policies. The proposed algorithm uses a risk function based on how far a state is from the state space known by the agent, that allows identifying and preventing adversarial situations. The success of the proposed defense has been evaluated in 4 Atari games. [ABSTRACT FROM AUTHOR]

Although there exist several approaches to the automatic colorization of the natural scene images and movies the problem of cartoon colorization has barely been considered. Therefore, this paper proposes a fully automatic pipeline to create a plausible colorization of vintage cartoons which is visually appealing to a human observer. Particularly, the Deep Cartoon Colorizer is proposed. The method incorporates an encoder–decoder convolutional neural network which is trained to map colors onto consecutive cartoon frames. The method was trained with 4944 images and tested on 34 vintage Disney cartoons including both the decolorized and the originally monochrome movies. In total 265388 cartoon frames were assessed including 246591 decolorized frames and 18797 originally monochrome ones. The resulting automatic colorizations were evaluated both quantitatively (by means of popular image perceptual quality measures) and qualitatively (by performing a human perception test). The resulting median values of the image quality measures range from 0.66 for Visual Information Fidelity and 0.85 for HaarPSI, through 0.94 for Structural Similarity Index, to 0.97 for Universal Image Quality Index which is a very good result. The subjective scores assigned by the human raters' on a ten-degree scale are on average equal to 6.11 and 6.63 for decolorized and monochrome frames respectively. This result also confirms that the desired effect of plausible colorization was obtained by the introduced approach. • The Deep Cartoon Colorizer is proposed for an automatic colorization of cartoons • The method is dedicated to old vintage cartoon episodes • The method uses an encoder–decoder convolutional neural network • The method was tested on a dataset containing 34 cartoon episodes • The colorization is plausible with SSIM equal to 0.94 and high raters' scores [ABSTRACT FROM AUTHOR]

Abstract The railway track fasteners play a critical role in fixing the track on the ballast bed. Achieving full automation of the fastener defect detection is significant in terms of ensuring track safety, and reducing maintains cost. In this paper, innovative and intelligent methods using image processing technologies and deep learning networks are proposed. In the first part, the traditional fastener positioning method based on image processing is reconsidered. In addition, a novel fastener defect detection and identification method using Dense-SIFT features is proposed which can achieve a better performance than the methods available in the literature. In the second part, VGG16 is trained for fastener defect detection and recognition. The result demonstrates that it is possible to carry out the defect detection of fasteners with CNN. Finally, Faster R-CNN is used for fastener defect detection to advance detection rate and efficiency. The fastener positioning and recognition can be carried out simultaneously. The time for the defect detection and classification is only one-tenth of the other methods mentioned above. Highlights • The Dense-SIFT, spatial pyramid decomposition and BOVW techniques are applied to the classification and this method achieves classification accuracy of 99.26%. • The DCNN method uses only one network for fastener defects detection and achieves a sound classification accuracy of 97.14%. • Faster R-CNN based method realizes the positioning and classification of the fasteners simultaneously and the time spent by Faster R-CNN is only about 10% of the time spent by the other methods. [ABSTRACT FROM AUTHOR]

Abstract Recently, automatic feature extraction and selection from unlabeled images that contain irrelevant patterns have been a proceeding interest. In this paper, an enhanced high-order Boltzmann machine is designed to promote the capacity of feature extraction and selection in a unified context. First, gating mechanism is employed for feature selection in comparison with conventional approaches. Then, two sets of hidden variables that the one set is real-valued latent variables and the other is spike latent variables are introduced to model the covariance structure of local patches, which can boost the abilities of feature learning and feature selection in turn. Simultaneously, the proposed model can infer in parallel via easy block Gibbs sampling without much training difficulty. Last, several extensions of the proposed model are developed to cope with different scenes. The massive performances obtained from various visual tasks have demonstrated that the proposed model can reach the highly improved performances over several currently excellent methods. [ABSTRACT FROM AUTHOR]

Abstract This paper concerns the issue of monitoring elderly behavior in the context of ambient assisted living (AAL). Under the framework of online daily habit modeling (ODHM), we employ the emergent representation for activities of daily living (ADLs) with position-based stigmergy, and then combine it with convolution neural networks (CNNs) to accomplish the tasks of recognizing ADLs. In addition, we propose a new paradigm of activity summarization with the robustness to break interruptions. Radio tomographic imaging (RTI) is promoted as a simple yet flexible way of facilitating the required position-based stigmergy. Such position-based AAL systems can benefit the advantages of having no need any sophisticated domain models in analyzing and understanding ADLs while no burden training is involved in ODHM. Moreover, the emergent based data aggregation and deep learning of CNN together allow the recognition of ADLs at a fine-grained level, which contributes to the performance improvement of ODHM. Experimental results demonstrate the effectiveness of the proposed approach. [ABSTRACT FROM AUTHOR]

This paper describes a novel DNN-based system, named PD3net, that detects multiple people from a single depth image, in real time. The proposed neural network processes a depth image and outputs a likelihood map in image coordinates, where each detection corresponds to a Gaussian-shaped local distribution, centered at each person's head. This likelihood map encodes both the number of detected people as well as their position in the image, from which the 3D position can be computed. The proposed DNN includes spatially separated convolutions to increase performance, and runs in real-time with low budget GPUs. We use synthetic data for initially training the network, followed by fine tuning with a small amount of real data. This allows adapting the network to different scenarios without needing large and manually labeled image datasets. Due to that, the people detection system presented in this paper has numerous potential applications in different fields, such as capacity control, automatic video-surveillance, people or groups behavior analysis, healthcare or monitoring and assistance of elderly people in ambient assisted living environments. In addition, the use of depth information does not allow recognizing the identity of people in the scene, thus enabling their detection while preserving their privacy. The proposed DNN has been experimentally evaluated and compared with other state-of-the-art approaches, including both classical and DNN-based solutions, under a wide range of experimental conditions. The achieved results allows concluding that the proposed architecture and the training strategy are effective, and the network generalize to work with scenes different from those used during training. We also demonstrate that our proposal outperforms existing methods and can accurately detect people in scenes with significant occlusions. • Robust system to detect people only using depth information from a depth camera. • System outperforms state-of-the-art methods in different datasets without fine-tuning. • Proposal runs in real time using conventional GPUs. • Computational demands are independent of the number of people in the scene. • Generated database is available to the research community. [ABSTRACT FROM AUTHOR]

In this fast-moving digital era, millions of images are added to repositories every millisecond. These images are context-rich images with ample underlying data that are extracted and analyzed by deep neural networks for innumerable research purposes. Significant breakthroughs in deep neural networks are harvesting multi-fold benefits and reaching higher accuracy and prediction rates. Though there is vertical progress in deep learning models there is always a threat to security which disintegrates the strength of deep neural networks. Archived images that an adversary manipulates will still look indistinguishable to the user. Such manipulations called adversarial attacks that trick deep learning models into making wrong predictions are on the rise nowadays. This paper reviews the varied types of adversarial attacks shedding light on their origin, their progressive advancement in multidimensions, along with their applications and challenges. [ABSTRACT FROM AUTHOR]

Anomaly detection in industrial environments aims at detecting anomalies in the monitoring data of industrial machinery or equipment, as soon as possible, preferably presenting real-time alarms, to alert the monitoring staff and start maintenance activities timely. In this paper, the problem of anomaly detection of an industrial furnace is tackled, for the real-time recognition of punctual anomalies on multivariate time series. To this aim, a real-time anomaly detection approach is proposed: first, time series acquired from the real machinery are filtered, to select those of interest for possible anomalies, and pre-processed, to obtain sliding windows for real-time detection, then distinct univariate models are applied, to identify different anomaly types. For the application considered here, data regarding the machinery behaviour were available only for normal functioning, thus an unsupervised approach is chosen. In particular, deep learning models based on autoencoders are used to detect punctual anomalies, by reconstructing each window and evaluating the reconstruction error of its last point. An extensive set of autoencoder models is proposed, with varying architecture in terms of type of model (vanilla/variational autoencoders), type of layers (fully connected/LSTM/BiLSTM), and hyperparameters (number of layers, intermediate sizes, BiLSTM type). Available data are split, and used to train the models, and to test them on the normal signal and on synthetic anomalies injected on it, which are of particular interest and were designed according to domain experts. Performances of the proposed models show differences among them, depending on the model architecture. The most efficient models, in terms of F1 score of detection and number of parameters, are identified by their t-test comparison, and the capability of detecting anomalies online is demonstrated. In particular, the proposed anomaly detection approach, including a selected autoencoder with LSTM layers, is able to correctly recognize normal trends, with very few false positives, and promptly give alarms as different anomalous trends appear. [Display omitted] • Definition and implementation of an online Anomaly Detection approach. • Comparison of different architectures of the autoencoder-based models. • Correct recognition of normal signal and detection of synthetic anomalies. • Application for industrial vacuum furnace predictive maintenance. [ABSTRACT FROM AUTHOR]

This paper proposes an intelligent text classification framework for a resource-constrained language like Bengali, which is considered a challenging task due to the lack of standard corpora, appropriate hyper-parameter tuning method, and pre-trained language-specific embedding. The proposed framework comprises an average meta-embedding feature fusion module and a convolutions neural network module called AVG-M+CNN. This work also proposes an algorithm, i.e., automatic hyperparameter tuning and selection, for enhancing the performance of the AVG-M+CNN technique. All meta-embedding models are evaluated using the intrinsic , e.g., semantic, syntactic, relatedness word similarity, analogy tasks and extrinsic evaluators. The intrinsic evaluator evaluates 200 Bengali semantic, syntactic and relatedness word pairs. Spearman (ρ ˆ), Pearson (r ˆ) and cosine similarity correlations are used to evaluate 18 individual embedding and 9 meta-embedding models. The 3COSADD and 3COSMUL evaluators evaluate the 300 analogy tasks. The extrinsic evaluator evaluates a total of 156 classification models on four corpora: BARD, IndicNLP, Prothom-Alo and B T C C 11 (a newly developed corpus having eleven distinct categories). Among these, the AVG-M+CNN model achieves the highest accuracy regarding four Bengali corpora: 95.92 ±.001% for BARD, 93.10 ±.001% for Prothom-Alo, 90.07 ±.001% for B T C C 11 and 87.44 ±.001% for IndicNLP, respectively. [ABSTRACT FROM AUTHOR]

Pests are diverse and the available datasets often contain an uneven number of examples for different pests (a.k.a., the long-tail distribution). This poses a great challenge to learning-based classification methods, especially deep networks, and often leads to degraded performance, especially for the minority (tail) classes. This paper presents a deep learning integration architecture based on decoupling training and fusion learning, which integrates different models with complementary performance on pest datasets with a long-tailed distribution to improve the overall classification performance of pests. A deep neural network is designed that fuses two complementary deep learning models at the feature level, which consists of a convolution neural network (ConvNeXt) and a Swin Transformer model for decoupling training. Experiments are conducted using three datasets (d0, insect, and IP102), and evaluation on accuracy, recall, and F1-Score is reported. For the large-scale pest dataset with long-tailed distribution IP102, the accuracy achieves 76.1%, which outperforms the state-of-the-art methods. In addition, the accuracy for d0 and insect datasets are 98.5% and 92.3%, respectively. [ABSTRACT FROM AUTHOR]

Single image super-resolution (SISR) aims to recover clear high-resolution images from low-resolution images, which has made great progress with the development of deep learning these years. Scene text image super-resolution (STISR) is a subfield of SISR with the goal of increasing the resolution of a low-resolution text image and enhancing the readability of characters in the image. Despite significant improvements in recent approaches, STISR remains a challenging task due to the diversity of background, text appearances and layouts, etc. This paper presents a Perceiving Multiple Representations (PerMR) method for better super-resolution performances in scene text images. PerMR is a unified network that combines super-resolution with text recognition and exploits the recognizer's feedback to facilitate super-resolution. Specifically, contextual information from the text decoder is extracted to provide sequence-specific guidance and enable the super-resolution model to pay more attention to the text region. Meanwhile, low-level and high-level visual features from the vision backbone of the recognition network are integrated to further improve visual quality. Additionally, we incorporate a frequency branch into the vanilla convolution unit, which efficiently enhances global and local feature representations. Experiments on the STISR benchmark dataset TextZoom validate that PerMR can not only generate more distinguishable images, but also outperforms the current state-of-the-art methods. PerMR boosts the average recognition accuracy by 5.9% using ASTER, 5.8% using MORAN and 10.6% using CRNN compared to the baseline model TSRN. PerMR outperforms the advanced method TPGSR-3 by 1.4% on ASTER, 0.1% on MORAN, 0.2% on CRNN and boosts TATT by 0.6% on ASTER and 1.1% on MORAN respectively. Furthermore, PerMR demonstrates good robustness and generalization when tackling low-quality text images in multiple scene text recognition datasets. The experiment results verify the capabilities of PerMR to boost text recognition performance. [ABSTRACT FROM AUTHOR]

The electrocardiogram (ECG) is a standard test to monitor the activity of the heart. Many cardiac abnormalities are manifested in the ECG including arrhythmia that refers to an abnormal heart rhythm. The basis of arrhythmia diagnosis is the identification of normal versus abnormal heart beats, and their correct classification based on ECG morphology. This paper proposes a novel and robust approach for representation learning of ECG sequences using a LSTM autoencoder for anomaly detection. The encoder part encodes the ECG signal into a lower dimensional latent space representation and decoder part then tries to reconstruct the specified ECG signal. The model is trained only on normal (non-anomalous) ECG signals. Then reconstruction loss of test ECG signals are calculated. Next a reconstruction loss threshold value is determined from the frequency distribution of the reconstruction losses so that from the reconstruction loss value above a certain threshold is determined as anomaly, otherwise it will be treated as normal. Determination of threshold is done using manual and Kapur's automated thresholding procedures. The aforementioned model has been applied on publicly available ECG5000 dataset. From the experimental results it is observed that the proposed model achieved more than 98% accuracy having precision, recall and F1 values more than 0.94, 0.97, 0.96 respectively. The performance of the proposed method is also found to be superior in most of the cases as compared to the results of seven other recent counter-part methods reported in the literature. [ABSTRACT FROM AUTHOR]

Large-scale model sizes and expensive computing costs cause the challenge of deploying and applying large pre-trained models. Hence, this paper presents a novel Triple Concepts Attention Mechanism and a lightweight TCAMixer model for edge devices to classify texts. Furthermore, the TCAMixer abstracts textual concepts in a human way, which is unmatched by other counterparts such as pNLP-Mixer (a projection-based MLP-Mixer model for Nature Language Processing) and HyperMixer (a hyper network using dynamic token-mixing layers). Experimental results on several public datasets demonstrate that the TCAMixer outperforms the counterparts by a significant margin, for example, achieving 3% higher accuracy with a smaller model size of 0.177M. Additionally, the TCAMixer achieves a performance of 85% to 98.7% compared to that of large pre-trained models but only occupies 1/3000 to 1/2000 of their size on most test datasets. [ABSTRACT FROM AUTHOR]

Meter error is one of the main contributing factors to unexpected fuel losses or gains in storage tanks at service stations. Although fuel dispensers are expected to be calibrated to standard accuracy periodically to ensure fair and reliable trade in the fuel market, some fuel retailers are unable to keep up with the standards. The current industry practice relies on onsite inspection to identify the issue, which leads to a cost burden if inspections are scheduled too frequently. To the best of our knowledge, there is no previously reported research tailored to the remote meter error detection problem. In this paper, we propose a novel framework for remote and automatic meter error detection via a data-driven approach based on inventory data and fuel transaction data. Specifically, we propose to use mean shift change point detection methods, including statistical-based as well as deep learning-based methods (LSTM-VAE, VAE, Kernel learning), to approach the problem. We present results on our data sets containing both real-world and simulated meter error data, and further evaluate these methods on several widely-used benchmark datasets, to assess their validity, advantages and limitations. The obtained results show that LSTM-VAE outperforms other models in most of the settings for the meter error dataset and the benchmark datasets. [ABSTRACT FROM AUTHOR]

Grape leaf diseases seriously affect grape yield and directly lead to the loss of the industrial economy. An accurate diagnosis is crucial for effective spraying and control. Presently, the combination of image blending and a deep recognition model is the most effective method for accurately identifying grape leaf diseases. Existing methods of image blending depend primarily on single-channel and multi-scale channel blending, computing and optimizing the Poisson equation, and the iterative optimization of learning-based methods for simple composite results. Although these methods have had some success, they are not able to simultaneously address lesion and leaf gap size, color bleeding, and edge smoothness. In this paper, we propose a novel image blending method named cascade wavelet attention network (CWAN). Specifically, CWAN consists of a cascade wavelet attention (CWA) and a blending autoencoder. First, we build a CWA to decompose the image into a multi-level and multi-scale frequency domain image. Then, the detailed features of the source and target domains are extracted, and the edges are smoothed. Following this, a shared decoder is designed to fuse these features in the corresponding frequency-domain layers, and the result serves as the attention feature. Finally, we design a blending autoencoder with a feature fusion mechanism to seamlessly blend lesions and leaves enhanced by CWA. We also construct a weighted blended loss function to avoid leaf color infiltration into the lesion in the blended image. Comprehensive experiments on two grape leaf disease datasets. Qualitative and quantitative results demonstrate the superiority and effectiveness of our model. • A novel method called CWAN for data enhancement of grape leaf disease is proposed. • The CWA is designed to enhance the textures and smooth the edge of blending. • The blending autoencoder with feature fusion mechanism can seamlessly blend images. • An unpaired image composition dataset is created for grape leaf disease blending. • The image shows superior performance on the texture and seam transition. [ABSTRACT FROM AUTHOR]

Inspired by heterogeneity of rapid-increasing 3D pavement data and the generalization ability of transfer learning, a robust and generalized framework for cross-scene 3D pavement-crack detection and attribute extraction was proposed in this paper, called profile component decomposition model with holistically nested edge detection (PCDM-HED). The core purpose of the PCDM-HED is to construct the enhanced deep edge features. By applying profile frequency, sparse characteristics of 3D profiles, and fusing the multiscale and multilevel edge characteristics of 3D depth maps in HED network, the robust enhanced feature can highlight the essential properties of cracks in heterogeneous 3D data. It overcomes the complex domain shifts caused by different 3D imaging conditions, pavement textures, and crack distributions in heterogeneous data. Cross-domain transfer experiments were carried out over seven 3D/2D datasets with 915 pavement sections. The results show that proposed PCDM-HED achieved average buffered Hausdorff scores of 90.17 to 96.42, recall scores of 0.84 to 0.91, and F-values of 0.85 to 0.89 in six different datasets without labeled samples. Compared with 9 groups of comparison results, including the traditional and related state-of-the-art methods, the transfer generalization effect of proposed PCDM-HED is more than 23% higher than that of comparison results. The proposed PCDM-HED makes full use of limited off-the-shelf samples, demonstrated strong transfer learning capability. It provides an effective solution for heterogeneous 3D pavement-crack detection tasks in engineering, in the case of limited labeled samples or even no corresponding labeled samples. • A 3D pavement crack detection with generalized transfer learning ability is proposed. • An enhanced deep edge feature is developed for heterogeneous 3D pavement data. • The PCDM-HED achieves BHS of 90.17 to 96.42 in heterogeneous without labeled samples. • The PCDM-HED demonstrates generalization ability across different 3D pavement scenes. [ABSTRACT FROM AUTHOR]

Detecting crops accurately is the key challenge for harvesting robots, and deep learning methods are commonly used for this purpose. However, these methods require a large amount of training data, which can be a constraint. Data augmentation is frequently recommended to enlarge the size of training data and enhance the performance of the detection models. In this paper, we propose the use of RandAugment (RA) to improve crop detection performance by applying geometric, photometric, and partial occlusion transformations. We experimented with different transformation lists using popular detection networks on Tomato datasets. YOLOv3 with geometric transformations and partial occlusion achieved the highest accuracy of 76.42, showing a 3.47 increase over the baseline model without augmentation. RA also improved detection accuracy on other public datasets such as Apple, Kiwi, and Mango. We found that the optimal combination of transformations varied for each network to achieve the best performance. Additionally, we implemented YOLOv3 with the optimal transformation list to demonstrate the robotic tomato harvesting system and achieved a computing time of less than 60 ms from capturing the camera image to receiving the detection results. [ABSTRACT FROM AUTHOR]

In the context of mechanical engineering design, the field of machine learning accelerated topology optimization is dominated by image-based models trained in a supervised manner on datasets with a limited diversity of boundary conditions. State-of-the-art methods show poor generalization capabilities and are strongly coupled to finite element mesh resolution, hindering scalability. In this paper, we leverage the explicit topology parameterization of the moving morphable components (MMC) framework to train a deep learning model that directly generates geometric design variables using a model architecture that is independent of the finite element mesh used for structural analysis. The developed model is trained on a large dataset of boundary conditions. Despite achieving state-of-the-art regression loss, evaluations reveal that direct-design approaches generate topologies with poor mechanical performance. Specifically, the model-generated topologies have, on average, a stiffness 11.48% lower than conventional MMC designs, as evidenced by in-distribution and out-of-distribution test samples. We demonstrate that this is due to the incompatibility between the regression loss function typically used in literature and the topology optimization objective of compliance minimization. To address this issue, we propose a novel acceleration approach that leverages the trained model to generate improved initial designs for conventional optimization. Specifically, the deep learning model is used to generate an initial design, which is then refined by conventional optimization to arrive at a final, optimal design. This approach shows a computation time-saving of 36.84% without sacrificing the final mechanical performance of the optimal topology compared to conventional optimization starting from a uniform initial layout. [ABSTRACT FROM AUTHOR]

Fetal heart monitoring and early disease detection using non-invasive fetal electrocardiograms (fECG) can help substantially to reduce infant death through improved diagnosis of Coronary Heart Disease (CHD) in the fetus. Despite the potential benefits, non-invasive fECG extraction from maternal abdominal ECG (mECG) is a challenging problem due to multiple factors such as the overlap of maternal and fetal R-peaks, low amplitude of fECG, and various systematic and environmental noises. Conventional fECG extraction techniques, such as adaptive filters, independent component analysis (ICA), empirical mode decomposition (EMD), etc., face various performance issues due to the fECG extraction challenges. In this paper, we proposed a novel deep learning architecture, LinkNet++ (motivated by the original LinkNet) to extract fECG from abdominal mECG automatically and efficiently using two different publicly available datasets. LinkNet++ is equipped with a feature-addition method to combine deep and shallow levels with residual blocks to overcome the limitations of U-Net and UNet++ models. It also has deep supervised and densely connected convolution blocks to overcome the limitations of the original LinkNet. The proposed LinkNet++ model was evaluated using fECG signal reconstruction and fetal QRS (fQRS) detection. As a signal-to-signal synthesis model, LinkNet++ performed very well in two real-life datasets and achieved 85.58% and 87.60% Pearson correlation coefficients (PCC) between the ground truth and predicted fECG on two datasets, respectively. In terms of fQRS detection, it also outperformed most of the previous works and showed excellent performance with more than 99% of F1 scores on both datasets. Our results indicate that the proposed model can potentially extract fECG non-invasively with excellent signal quality, thereby providing an excellent diagnostic tool for various fetal heart diseases. • A novel LinkNet++ with deep supervision and dense blocks is proposed for automatic extraction of fECG from mECG. • A rigorous leave one subject out cross-validation study is conducted using two datasets to build a generalizable model. • The innovative LinkNet++ model outperforms all previous models and standard signal processing methodologies. [ABSTRACT FROM AUTHOR]

Even though deep learning models can be applied for the surface defect detection of industry products, it remains a challenge to widely deploy them on embedded carriers due to high computation and large memory requirements. To deal with this challenge, this paper proposed to search for the best pruning architecture of the YOLO(You Only Look Once)v3 model with the Particle Swarm Optimization evolutionary algorithm for model channel Pruning (PSOP), in addition to focusing on the importance of filters in a convolutional layer. An automatic sparsity training technique is induced in PSOP to generate an initial model before pruning. PSO searching speed is accelerated as a fast evaluation method is employed during training. A linear multi-object strategy is utilized to help PSO find the best structure for the pruning model in the searching period. PSOP outperforms many outstanding pruning methods. For instance, evaluated by our copper elbow dataset, the pruned YOLOv3 saves 62.93% of the model volume, 72.28% floating-point operations, and 62.96% parameter size with only 3.13% accuracy loss. The pruned YOLOv3 model is successfully deployed on NVIDIA Jetson Nano B01 embedded platform to improve copper elbow plant productivity. [ABSTRACT FROM AUTHOR]

Level Crossings are among the most critical railway assets, concerning both the risk of accidents and their maintainability, due to intersections with promiscuous traffic and difficulties in remotely monitoring their health status. Failures can be originated from several factors, including malfunctions in the bar mechanisms and warning devices, such as light signals and bells. This paper focuses on the intelligent detection of anomalies in warning bells through non-intrusive acoustic monitoring by: (1) introducing a new concept for autonomous monitoring of level crossings; (2) generating and sharing a specific dataset collecting relevant audio signals from publicly available audio recordings; (3) implementing and evaluating a solution combining deep learning and transfer learning for warning bell detection. The results show a high accuracy in detecting anomalies and suggest viability of the approach in real-world applications, especially where network cameras with on-board microphones are installed for multi-purpose level crossing surveillance. • New concept for autonomous monitoring of level crossings through audio–video analytics. • Novel approach for level crossing bell detection using deep learning and transfer learning. • Generation of a specific dataset collecting relevant audio signals. • Results show high accuracy and suggest feasibility within smart railway maintenance. [ABSTRACT FROM AUTHOR]

Due to the remarkable ability of Deep learning (DL) to abstract hidden information, it has been proven to be a powerful tool in many tasks related to the design of metamaterials. DL-aided design techniques can be generally categorized into two types, including forward designs, which are training surrogate models to accelerate the simulation process, and inverse design, which uses inverse modeling techniques to generate the design that satisfies the corresponding requirement. Although generative models have a unique capability to generate multiple designs instantly with random information, they often underperform in accuracy compared to designs based on optimization techniques. In this paper, a hybrid design framework combining the advantages of both DL forward design and the inverse design based on the mixture density network (MDN) is proposed. Then the proposed framework is implemented for the inverse design of S-shaped perforated auxetic metamaterial. The hybrid design framework inherited the one-to-many mapping capability of MDN and has great capability of generating designs with designated mechanical properties at less than 10% relative errors, in most design scenarios (over 95% in the test set), at one to two orders of magnitude less computational cost compared to optimization-based forward design. • A hybrid method with MDN and DNN generates designs at low cost but good accuracy. • A case study of auxetic metamaterial shows the effectiveness of the hybrid method. • The comparison with current DL design models shows the advantage of our model. [ABSTRACT FROM AUTHOR]

As the complexity of power systems increases, traditional model-driven methods for online frequency stability prediction (FSP) encounter constraints in both accuracy and efficiency. To enhance the accuracy and efficiency of FSP, an data-driven method using CoAtNet and SHAP values is proposed. By leveraging the combination of convolution and attention mechanisms, CoAtNet addresses the limitation of traditional deep learning approaches that may not be able to extract data features comprehensively. Moreover, selecting all features as input into a deep-learning model may cause a substantial computation burden. It is thus impractical for CoAtNet to perform FSP of large-scale power systems. For this problem, this paper develops a SHAP values-based feature selection method to select the effective features as input. This process greatly reduces the numerical complexity, maintaining a high prediction performance. Additionally, the marginally stable situation of the system frequency is ignored by most researchers. A frequency security index to identify marginally stable situations is thus employed to generate the data labels, which are classed as "absolute security", "relative security", and "insecurity". Finally, verified by the comparison simulation, the proposed model outperforms other models with accuracies of 98.80% on the modified IEEE 39-bus system and 99.04% on the modified ACTIVSg500 system. [ABSTRACT FROM AUTHOR]

Supercapacitors have been widely used in many fields. The safe and stable operation of supercapacitors requires accurate remaining useful life (RUL) prediction. This paper proposes a two-stage online RUL prediction framework based on the bidirectional long short-term memory (BiLSTM) network and the H ∞ observer. In Stage 1, the BiLSTM network as well as the Bayesian Optimization algorithm is used to estimate capacitance. The Bayesian Optimization (BO) algorithm aims to optimize the hyperparameters of the BiLSTM network. In Stage 2, the estimated capacitance is handled by the moving average filter (MAF) to alleviate short-term fluctuations. The double exponential model is employed to describe the degradation trajectory, and the model is iteratively updated by the H ∞ observer with the estimated capacitance as the measurement. Experiments are implemented to verify the validation of the proposed framework. The results indicate that the BiLSTM network can explain more than 99% variation of capacitance. The deviation of RUL prediction is less than 6% in most cases, and it is even less than 2.5% in late phase. The proposed framework takes the advantages of online deployment, and achieves competitive prediction accuracy when compared with offline methods. • An online RUL prediction framework is proposed based on the H ∞ observer and the BiLSTM. • A new input for the BiLSTM (i.e. difference voltage) is defined based on data analysis. • BO algorithm is introduced to optimize the hyperparameters of the BiLSTM network. • MAF is used to alleviate data fluctuations and smooth the predicted capacitance. [ABSTRACT FROM AUTHOR]

In recent research, problems with biased datasets or domain shift have presented challenges to the practical applications of deep learning methods. In this paper, we propose a simple method using adversarial learning combined with contrastive learning and domain adaptation to solve the domain-shift problem. Because domain shift is caused by differences in the distribution of data across domains, different approaches have been proposed to resolve this by rendering the data distributions closer through adversarial training. Contrastive learning is a method of acquiring valid feature representations for downstream tasks. By combining these approaches, we aim to achieve better domain adaptation. The proposed method is simple and intuitive. By introducing a domain discriminator into SimCLR, which is a typical contrastive learning model, and training it in an adversarial manner, the feature vectors of the source and target domains are rendered closer to acquire domain-invariant features. The proposed approach facilitates high-performance pretraining without labels and demonstrates a significant improvement in accuracy in comparison to standard benchmark methods, including conventional supervised models and SimCLR. • We investigated an unsupervised domain adaptation task. • Combining adversarial and contrastive learning can address the domain-shift issue. • The proposed method is practical and implementation is very simple. [ABSTRACT FROM AUTHOR]

Home automation is seen as a potential pillar of the smart city revolution that combines smart mobility, lifestyle and ecosystem governed by intelligent sensors connected to the internet. Households can save money and be more comfortable with automated appliances. The cost of electricity and user comfort are fundamentally contradictory, so they can be presented as a dynamic multi-objective optimization problem with fluctuating priorities for the customer to use various devices at different times. For this reason, this paper proposes an advanced Intelligent Home Energy Management (IHEM) approach based on reinforcement learning to achieve home demand response (DR) efficiency. The optimal formulation of the one-hour-ahead energy consumption scheduling problem is considered a Markov Decision Process (MDP) with discrete time steps. An efficient Neural Network (NN)-based approach with a Q-learning algorithm is developed to address this problem, enabling the IHEM system to achieve better cost-effective scheduling performance. The accurate data of electricity price and energy supplied by the Photovoltaic (PV) system are analyzed in sliding periods by machine learning for uncertainty prediction. Using the newly developed approach, which has the dual objective of minimizing the electricity bill, it is possible to obtain scheduling decisions for appliances and energy storage. The results show that the proposed optimization method reduces the monthly electricity costs by 20% compared to the Integer Linear Programming (ILP)-based HEMS method. • Artificial intelligence (AI) approach to monitor household energy consumption. • An efficient deep reinforcement learning algorithm to control activity recognition in smart homes. • Design and implementation of Intelligent home energy management system. • Converge promptly and considerably reduce operating expenses adopting AI-based approach. [ABSTRACT FROM AUTHOR]

Temperature field prediction is of great importance in the thermal design of systems engineering, and building a surrogate model is an effective method for the task. Ensuring a high prediction performance for the surrogate models, especially deep learning models with high representational power and numerous parameters, typically requires a significant amount of labeled data. However, obtaining labeled data, particularly high-fidelity data can be prohibitively expensive. To solve this problem, this paper proposes a novel deep multi-fidelity modeling method for temperature field prediction, which takes advantage of low-fidelity data to boost performance with less high-fidelity data. First, a pithy pre-train and fine-tune paradigm is proposed for constructing the deep multi-fidelity model, which is straightforward and efficient, allowing for the effective utilization of information from various fidelity levels. Then, a physics-driven self-supervised learning method is proposed to learn the deep multi-fidelity model, which fully utilizes the physics characteristics of the heat transfer system and further reduces the dependence on large amounts of labeled low-fidelity data in the training process. Two diverse temperature field prediction problems are presented to validate the effectiveness of the proposed method. The results show that our approach can significantly improve the model's accuracy, reducing the required high-fidelity data for model construction. [ABSTRACT FROM AUTHOR]

Circular ribonucleic acids (circRNAs) are widely expressed in cells and tissues and play vital roles in cellular physiological processes. Their expressions are associated with clinicopathological features in cancer patients. Thus, they act as molecular biomarkers for tumor diagnosis, non-invasive monitoring, prognosis, and therapeutic intervention. Recent research has shown that circRNAs can interact with RNA-binding proteins (RBPs), which is a critical aspect for understanding circRNA functions. In this paper, we review the state-of-the-art deep learning and ensemble deep learning methods highlighting their strengths and weaknesses in circRNA-RBP interaction prediction. We further discuss new strategies for improving the existing methods. The existing circRNA-RBP interaction prediction methods are further classified as deep learning or ensemble deep learning. Moreover, we elaborate on the critical factors for cicRNA-RBP interactions, which can help the development of prediction models by providing necessary clarifications. This review further presents the benefits of using ensemble deep learning methods over single deep learning methods. Prediction performance improvements of ensemble deep learning methods over single deep learning methods are observed and the reasons for those improvements are discussed. Furthermore, this review discusses open problems of this research field and provides recommendations on future research directions. [ABSTRACT FROM AUTHOR]

Change detection is important in remote sensing image analysis. In recent years, significant breakthroughs have been made in change detection algorithms based on deep learning. However, due to continuous downsampling, the detection results of these algorithms still have serious detection errors, detection omissions and edge blurring. Aiming at these problems, this paper proposes a dual-branch network for change detection. The network has two branches, which are used to extract the depth-variant semantic features of the multi-temporal image pairs and the respective features of each image respectively. In addition, we designed a Multi-scale Strip Convolution Module (MSCM) to extract the multi-scale features of the image, a new Spatial Attention Module (SAM) to strengthen the feature representation of changing regions, and a Feature Fusion Network (FFN) to guide the fusion between multiple features of the two branches. Experimental results show that the proposed method substantially mitigates detection errors, detection omissions and obtains sharper edges, it outperforms other current algorithms. [ABSTRACT FROM AUTHOR]

In order to solve the problems of low efficiency and dependability of human experience in shot hole depth recognition by manual interpretation of the massive drilling videos, this paper proposed a repetition counting framework from the perspective of representation object motion, and firstly studied and realized the automatic shot hole depth recognition by adopting the proposed framework by combining multiple deep learning-based models. Firstly, the periodic motion characteristics of the objects in the drilling video were analyzed, and the motion extraction model of the representative object was established based on YOLOv5 and Deep SORT algorithm to obtain the spatiotemporal sequence. Then, the sequence noise and redundant information were filtered according to the spatiotemporal relationship between the drill pipes and workers, to obtain a coherent, smooth, and clean motion sequence. On this basis, a drill pipe motion counting model was established by using the YOLOv5 network to detect the temporal boundary and repetition number of the drill pipe motion. Finally, combined with the pipe motion extraction model and counting model, the comprehensive reliable recognition method of shot hole depth was presented. The shot hole depth recognition trial was carried out on 1029 drilling videos from the Bureau of Geophysical Prospecting INC. The proposed method achieved 90.5% recognition accuracy (Acc) and 0.027 mean absolute error (MAE) with 64 frames per second (FPS) and 0.39 average recognition time ratio (R t). Therefore, the proposed repetition counting framework solved the problem of efficient and intelligent drilling depth recognition of massive drilling videos and thus provided a new idea for repetition counting. • A new repetition counting framework for shot hole depth recognition based on representative object motion was proposed. • The explicit features were added into the optimization of the framework components. • The multi-scale temporal correlation was transformed into process detection to handle the non-stationary period problem. • A sequence denoising method based on spatiotemporal relationship was proposed to handle the noise and redundant information. [ABSTRACT FROM AUTHOR]

Inverse kinematics is the basis for controlling the motion of robotic manipulators. It defines the required joint variables for the robotic end-effector accurately reach the desired location. Due to the derivation difficulty, computation complexity, singularity problem, and redundancy, analytical Inverse kinematics solutions pose numerous challenges to the operation of many robotic arms, especially for a manipulator with a high degree of freedom. This paper develops different Deep Learning networks for solving the Inverse kinematics problem of six- Degrees of Freedom robotic manipulators. The implemented neural architectures are Artificial Neural Network, Convolutional Neural Network, Long–Short Term Memory, Gated Recurrent Unit, and Bidirectional Long–Short Term Memory. In this context, we associate the proposed results with a specific tuning of Deep Learning network hyper-parameters (number of hidden layers, learning rate, Loss function, optimization algorithm, number of epochs, etc.). The Bidirectional Long–Short Term Memory network outperformed all proposed architectures. To be close as possible to the experimental results, we have included two types of noise in the training data set to validate which of the five proposed neural networks is more efficient. Furthermore, in this study, we compare the performance of analytical and soft computing solutions in generating robots' trajectories. We include this scenario, focusing on the advantage of implementing neural networks in avoiding the singularity problem that can occur using the analytical approach. In addition, we used the RoboDK simulator to show simulation results with real-world meaning. The performance of Deep Learning models depends on the complexity of the posed problem. Moreover, the complexity of the Inverse Kinematics problem is related to the number of Degrees of Freedom. At the end of this work, we evaluate the influence of the complexity of robotic manipulators on the proposed Deep Learning networks' performance. The results show that the implemented Deep Learning mechanisms performed well in reaching the desired pose of the end-effector. The proposed inverse kinematics strategies apply to other manipulators with different numbers of Degrees of Freedom. [ABSTRACT FROM AUTHOR]

Tumor lesion segmentation and staging in cancer patients are one of the most challenging tasks for radiologists to recommend better treatment planning like radiation therapy, personalized medicine, and surgery. Recently, Deep Learning (DL) has emerged as an assistive technology to help radiologists to characterize the biology of tumors and manage cancer patients. Positron Emission Tomography/Computed Tomography (PET/CT) multi-modality image-based tumor segmentation has gained tremendous attraction. However, the fusion of PET and CT information exposes numerous serious challenges including intra-class variability, contrast issues, modality discrepancy (difference in shape, and size of tumor), and the blurred boundaries between tumor and normal tissues (low specificity). To address these challenges, various DL-based tumor auto-segmentation methods have been proposed to consider complementary and contradictory anatomical and functional information of multi-modality PET/CT. This survey paper provides an in-depth exploration of these auto-segmentation methods. First, we discuss PET, CT weaknesses, the need for PET/CT, and the challenge of multi-modality PET/CT images. Second, we provide a detailed discussion of the parameters used to evaluate the achievements and limitations of the reviewed methods. Third, we classify the existing solutions into three major groups based on the model architecture design such as single network, multiple networks, and hybrid network models. The multiple networks are further divided into ensembles, multi-task, and Generative Adversarial Network (GAN) models. Furthermore, we present a discussion on these solutions to improve segmentation performance along with their strengths and weaknesses. Finally, we present a discussion on open research challenges and recommend potential future directions. • Provide a comprehensive review of bi-model PET/CT images for lesion segmentation. • Categorized the existing study into single, multiple, and hybrid network models. • Highlight and discuss the multiple networks into ensemble, multi-task, and GAN. • Outline comparative analysis of reviewed methods along with strengths and weaknesses. • Identify open research problems and recommend potential future research avenues. [ABSTRACT FROM AUTHOR]

Time series data are ubiquitous in human society and nature, and classification is one of the most significant problems in the field of time series mining. Although it has been intensively studied, and has achieved significant results and successful applications, it is still a challenging problem, which requires capturing of multi-scale features of one-dimensional or multi-dimensional time series in variable length. In this paper, we propose a novel time series feature extraction block named Convolutional Gated Linear Units (CGLU), which is a combination of convolutional operations and Gated Linear Units for adaptively extracting local temporal features of time series. Combined with a temporal maxpooling block, it can extract global temporal features. To capture more diverse features, the Inception architecture is adopted to organize the CGLUs with different convolution kernel sizes, which result in the Convolutional GLU network. In order to evaluate the performance, we conduct extensive experiments on the UCR time series datasets (one-dimension) and UEA datasets (multi-dimension). Compared with baselines, our model obtains best results in terms of classification accuracy and training speed, which demonstrate effectiveness and efficiency of CGLUs and Conv-GLU network on time series classification tasks. [ABSTRACT FROM AUTHOR]

This paper proposes a deep neural network-based topology optimization acceleration method for irregular design domains that predicts (near-)optimal topologies. A topology optimization problem is a complex non-Euclidean data, which can be embedded in a graph form, and a graph neural network encodes it to Euclidean data such as vectors and matrices. The encoded information is applied to a multi-layer perceptron-based implicit neural representation model, and the multi-layer perceptron approximator predicts the compliance optimal material distribution. The prediction performance of the proposed encoder-approximator architecture is evaluated for several topology optimization problems. The trained network provides 96.6% compliance accuracy, except for 8.0% of the outliers. The two criteria have been investigated to estimate potential outliers, and post-optimization can resolve the outlier within fewer iterations than the original optimization. • A deep learning model to predict near-optimal topology available for irregular domains. • Optimization problems are represented as graphs and processed by a graph neural network. • Distance and optimality measures are investigated for estimating potential outliers. • Post-optimization resolves outliers thereby achieving accelerated optimization. [ABSTRACT FROM AUTHOR]

Robot object detection is important for the realisation of robot intelligence. Currently, deep learning-based object detection algorithms are used for robotic object detection. However, it faces some challenges in practical applications, such as the fact that robots frequently use resource-constrained devices, resulting in detection algorithms with long computation times and undesired detection rates. In order to address these concerns, this paper proposes a lightweight object detection algorithm for robots with an improved YOLOv5. To reduce the amount of processing required for feature extraction and increase the speed of detection, the C3Ghost and GhostConv modules have been introduced into the YOLOv5 backbone. The DWConv module was used in conjunction with the C3Ghost module in the YOLOv5 neck network to further reduce the number of model parameters and maintain accuracy. The CA (Coordinated Attention) module is also introduced to improve the extraction of features from detected objects and suppress irrelevant features, thus improving the algorithm's detection accuracy. To verify the performance of the method, we tested it with a self-built dataset (4561 robot images in total) and the PascalVOC dataset respectively. The results show that compared with the YOLOv5s on the self-built dataset, the algorithm has a 54% decrease in FLOPs and a 52.53% decrease in the number of model parameters without a decrease in mAP (0.5). The effectiveness and superiority of the algorithm is demonstrated through case studies and comparisons. • Lightweight C3Ghost and GhostConv modules are introduced in the YOLOv5 backbone network to achieve model compression and maintain detection accuracy and speed. • C3Ghost and DWConv modules are introduced in YOLOv5 neck network to further reduce model parameters and improve the speed of feature fusion. • CA (Coordinated attention) module is also introduced to enhance the extraction of relevant features and suppress irrelevant features to improve the detection accuracy of the algorithm. • To demonstrate the algorithm's ability to solve real-world problems, we produced a dataset for the object detection task in the RoboMasterAI challenge hosted by DJI and experimentally verified that our proposed object detection algorithm is effective. [ABSTRACT FROM AUTHOR]

Video super-resolution methods focus on restoring high-resolution video frames from low-resolution videos with pre-defined degradations, and few consider compression. However, the videos on the Internet are compressed to reduce the massive size of video data. Thus, they also contain visual noises and artifacts due to the lossy compression. In this paper, we propose a video super-resolution model specific for compressed videos to suppress artifacts and restore high-resolution contents. We propose a two-stage super-resolution network for compressed videos that contains three new modules: cleanup and local alignment in the first stage, 3D-convolution temporal–spatial attention in the second stage. Specifically, the cleanup module is designed in stage-1 to suppress the artifacts and noises of low-quality inputs. The subsequent local alignment is designed to aggregate adjacent frames to maintain high-frequency details. The 3D-convolution attention is used in stage-2 to refine the weights of the feature maps. In addition, we propose the new CrfVideos dataset to facilitate fair comparisons, containing HEVC/H.265 compressed videos of different compression levels. Extensive experiments substantiate that our method outperforms state-of-the-art methods on the HEVC/H.265 compressed benchmark with medium to high compression rates by 0.3–0.7 dB in terms of PSNR and exceeds state-of-the-art methods on benchmark Vid4 by 0.2–0.8 dB. The code will be released at https://github.com/cvygkhv/decompression. [ABSTRACT FROM AUTHOR]

Detecting objects in hospital indoor environments is critical for scene understanding and can have various applications in healthcare. Deep learning algorithms have proven to be effective in object recognition from images or videos, but the availability of annotated datasets plays a crucial role in their successful application. However, there is a shortage of datasets for object detection in hospital settings, hindering the advancement of hospital indoor object detection algorithms. In this paper, we present the Hospital Indoor Object Detection (HIOD) dataset, consisting of 4,417 images covering 56 object categories. The HIOD dataset represents the frequently encountered objects in hospitals and comprises 51,869 annotated objects. The dataset is characterized by dense annotation, with an average of 11.7 objects and 6.8 object categories per image. An object detection benchmark was established using the HIOD dataset and eight state-of-the-art object detectors. The benchmark provides a comprehensive evaluation of the performance of the selected object detectors on a large and diverse set of images of objects commonly seen in hospital environments. The results of the benchmark can be used to compare and analyze the performance of different object detectors and identify their strengths and weaknesses for use in hospital environments. In the benchmark, one-stage detectors have shown superior performance compared to two-stage detectors of similar parameter sizes. In particular, YOLOv6-L was able to attain a mean Average Precision (mAP) of 51.7% while operating at a detection speed of 255 FPS. The benchmark and dataset can serve as a valuable resource for researchers and practitioners in the field of computer vision and robotics, helping to advance the development of more effective and efficient object detection algorithms for developing automated operations in hospitals such as robotic disinfection and patient assistance. • A new object detection dataset for artificial intelligence applications in hospitals. • A benchmark to compare and analyze the performance of different object detectors. • Improve object detection via hyperparameter optimization and transfer learning. • Evaluate the effects of image size and backbone on object detection performance. [ABSTRACT FROM AUTHOR]

Conventional intelligent fault diagnosis algorithms based on signal processing and pattern recognition have high demands on expert experience and poor generalization performance, which may not have good fault diagnosis performance in complex industrial fields. Meanwhile, the data acquisition system may suffer from cyber attacks when collecting vibration signals. The vibration signal has a very low signal-to-noise ratio (SNR), which seriously affects the accuracy of fault diagnosis. Aiming at the problem of fault diagnosis under low SNR, a new fault diagnosis framework based on a Multi-scale Split Dual Calibration Network with Periodic Information (PI-MSDCN) is proposed in this paper. In the fault diagnosis framework, a periodic block is constructed to automatically learn the periodic information of vibration signals through the neural network. The learned periodic information and raw vibration signal are used as the input data of MSDCN. Specifically, MSDCN uses convolution kernels of different sizes for different channels of input features to generate multi-scale features, and obtain mixed domain attention features for features with different scales respectively. Then, the attention feature is used as the threshold to remove the redundant information in the multi-scale feature adaptively. Finally, in order to calibrate the contribution of different scale features to fault diagnosis, the mixed-domain attention coefficients are applied to the corresponding features to obtain richer multi-scale attention features. The experimental studies under different levels of interference are performed to demonstrate the average accuracy of the proposed method is 92.91% (± 5. 08 %), which is superior to other existing results in literature. [Display omitted] • We propose a new period block into the diagnosis network that can automatically learn the period information from the adaptive fitting of different vibration signals through the neural network, which enhances the interpretability. • The developed MSDCAB realizes the removal of redundant information and dual calibration of attention weight for features with different scales. • The entire fault diagnosis framework is composed of periodic block and multiple MSDCABs, which can accurately predict the fault types of rolling bearings. [ABSTRACT FROM AUTHOR]

The leading causes of earthquakes are crustal movements, plate movements, and collisions. In recent years, many researchers in earthquake prediction have been predicting earthquakes from historical seismic data in local areas. This approach ignores the underlying internal patterns of crustal motion, plate movement, and collisions. This paper proposes a purely data-driven deep learning model called EPT. The model uses gated feature extraction blocks (GFEB) to mine potential crustal motion and plate movement patterns from global historical seismic catalog data. It uses them to aid mainshock prediction in each local, provincial region. Experiments show that this approach improves model prediction accuracy by up to 50 percent. We also use multi-headed self-attention for the first time to capture long-term dependencies within regional time series, highlighting links between focal features and compensating for the difficulty of focusing on longer-term information in long-term time series with long short-term memory networks (LSTM). In addition, we also use the gradient harmonization mechanism classification (GHMC) loss function for the first time in earthquake prediction, effectively addressing the problem of uneven data distribution across different earthquake magnitude ranges. Finally, we validated the effectiveness of the EPT model in five provincial datasets in mainland China, and the experimental results all achieved an accuracy of over 90 percent. [ABSTRACT FROM AUTHOR]

Sudden-onset disasters put forward new requirements for on the state authorities' ability to analyze public opinion sentiment. However, traditional sentiment analysis methods ignore the contextual semantic relationships and out-of-vocabulary words, and their computational resource utilization is excessive compared to their expected accuracy. In this paper, an ALBERT-based model combined with a text convolution neural network, a hierarchical attention mechanism and the latent Dirichlet allocation is proposed to create a hybrid model enhanced with topic knowledge for sentiment analysis of sudden-onset disasters. Weibo text data from a rainstorm disaster in China are used to evaluate the model's performance. Compared with the XLNet, DistilBERT and RoBERTa models, the experimental results demonstrate that the proposed approach is capable of achieving better performance by incorporating external topic knowledge into the language representation model to compensate for the limited vocabulary data. [ABSTRACT FROM AUTHOR]

Future industrial automation systems are anticipated to be shaped by intelligent technologies that allow for the adaptability of machines to the variations and uncertainties in processes and work environments. This paper is motivated by the need for devising new intelligent methods that enable efficient and scalable training of collaborative robots on a variety of tasks that foster their adaptability to new tasks and environments. Recent advances in deep Reinforcement Learning (RL) provide new possibilities to realize this vision. The state-of-the-art in deep RL offers proven algorithms that enable autonomous learning and mastery of a variety of robotic manipulation tasks with minimal human intervention. However, current deep RL algorithms predominantly specialize in a narrow range of tasks, are sample inefficient, and lack sufficient stability, which hinders their adoption in real-life, industrial settings. This paper develops and tests a Hyper-Actor Soft Actor–Critic (HASAC) deep RL framework based on the notions of task modularization and transfer learning to tackle this limitation. The goal of the proposed HASAC is to enhance an agent's adaptability to new tasks by transferring the learned policies of former tasks to the new task through a "hyper-actor". The HASAC framework is tested on the virtual robotic manipulation benchmark, Meta-World. Numerical experiments indicate superior performance by HASAC over state-of-the-art deep RL algorithms in terms of reward value, success rate, and task completion time. [ABSTRACT FROM AUTHOR]

Malware writers are usually focused on those platforms which are most used among common users, with the aim of attacking as many devices as possible. Due to this reason, Android has been heavily attacked for years. Efforts dedicated to combat Android malware are mainly concentrated on detection, in order to prevent malicious software to be installed in a target device. However, it is equally important to put effort into an automatic classification of the type, or family, of a malware sample, in order to establish which actions are necessary to mitigate the damage caused. In this paper, we present CANDYMAN, a tool that classifies Android malware families by combining dynamic analysis and Markov chains. A dynamic analysis process allows to extract representative information of a malware sample, in form of a sequence of states, while a Markov chain allows to model the transition probabilities between the states of the sequence, which will be used as features in the classification process. The space of features built is used to train classical Machine Learning, including methods for imbalanced learning, and Deep Learning algorithms, over a dataset of malware samples from different families, in order to evaluate the proposed method. Using a collection of 5,560 malware samples grouped into 179 different families (extracted from the Drebin dataset), and once made a selection based on a minimum number of relevant and valid samples, a final set of 4,442 samples grouped into 24 different malware families was used. The experimental results indicate a precision performance of 81.8% over this dataset. [ABSTRACT FROM AUTHOR]

To estimate the parameters of an induction motor in a data-based manner, this paper proposes a new offline method to estimate rotor resistance and excitation inductance based on the deep-Q-learning approach. In this method, parameter estimation can be facilitated without disturbing the model error or operating state. To achieve this goal, three important elements, namely observation, action, and reward, are appropriately designed. To improve the robustness and accelerate the convergence, a new concept, denoted as Q-sensitivity, is proposed and investigated in detail. The experimental results show that a high-Q-sensitivity design can allow the proposed method to obtain a fast and torque-maximized estimation. Results from the comparative studies confirm the accuracy and robustness of the proposed method. [ABSTRACT FROM AUTHOR]

After 2007–2008 crisis, it is clear that corporate credit scoring is becoming a key role in credit risk management. In this paper, we investigate the performances of credit scoring models applied to CDS data sets. The classification performance of deep learning algorithm such as deep belief networks with Restricted Boltzmann Machines are evaluated and compared with some popular credit scoring models such as logistic regression, multi-layer perceptron and support vector machine. The performance is assessed using the classification accuracy and the area under the receiver operating characteristic curve. It is found that DBN yields the best performance. [ABSTRACT FROM AUTHOR]

Obstructive sleep apnea (OSA) is a common type of sleep-related breathing disorder, and polysomnography (PSG) remains the gold standard for its diagnosis. However, it takes a significant amount of time to perform PSG in a well-equipped laboratory, and patients typically have to wait a long time for a PSG test. In view of this, over recent years portable and even wearable tools for OSA classification have been developed as a low-cost and easy-to-use alternative to PSG. In this paper, a deep neural network (DNN)-based model was developed to classify OSA severity using peripheral oxygen saturation (SpO 2) signals; it showed the following advantages. First, the presented model takes unsegmented SpO 2 signals recorded overnight as its input, and OSA severity is then classified as one of four levels as the output. Consequently, there is obviously no need to label segmented signals, and the tremendous amount of effort spent on signal segmentation and then annotation can be completely saved. Second, a high generalization ability is provided since the largest amount of data were used to test the model. This feature gives the model an improved reliability for clinical use. Notably, the outperformance of this work is highlighted in a two-level classification case (with a cutoff apnea–hypopnea index of 5), where the accuracy and the sensitivity increased to above 91% and 95%, respectively. [Display omitted] • A novel method is proposed to classify OSA severity using unsegmented SpO2 signals. • The time and effort spent on signal segmentation and annotation can be fully saved. • A high generalization ability is provided in the presented model. • This work significantly outperformed counterparts in terms of accuracy. [ABSTRACT FROM AUTHOR]