Search

There is increasing focus on applying deep learning methods to electrocardiograms (ECGs), with recent studies showing that neural networks (NNs) can predict future heart failure or atrial fibrillation from the ECG alone. However, large numbers of ECGs are needed to train NNs, and many ECGs are currently only in paper format, which are not suitable for NN training. We developed a fully-automated online ECG digitisation tool to convert scanned paper ECGs into digital signals. Using automated horizontal and vertical anchor point detection, the algorithm automatically segments the ECG image into separate images for the 12 leads and a dynamical morphological algorithm is then applied to extract the signal of interest. We then validated the performance of the algorithm on 515 digital ECGs, of which 45 were printed, scanned and redigitised. The automated digitisation tool achieved 99.0% correlation between the digitised signals and the ground truth ECG (n = 515 standard 3-by-4 ECGs) after excluding ECGs with overlap of lead signals. Without exclusion, the performance of average correlation was from 90 to 97% across the leads on all 3-by-4 ECGs. There was a 97% correlation for 12-by-1 and 3-by-1 ECG formats after excluding ECGs with overlap of lead signals. Without exclusion, the average correlation of some leads in 12-by-1 ECGs was 60–70% and the average correlation of 3-by-1 ECGs achieved 80–90%. ECGs that were printed, scanned, and redigitised, our tool achieved 96% correlation with the original signals. We have developed and validated a fully-automated, user-friendly, online ECG digitisation tool. Unlike other available tools, this does not require any manual segmentation of ECG signals. Our tool can facilitate the rapid and automated digitisation of large repositories of paper ECGs to allow them to be used for deep learning projects. [ABSTRACT FROM AUTHOR]

One of the most important strategies used to mitigate the adverse impacts of traffic growth on mobile networks is caching. By caching at the edge, the backhaul traffic load is reduced, and the quality of service for the user is increased. Developing an effective caching algorithm requires accurate prediction of the future popularity of the content, which is a challenging issue. In recent years, deep learning models have achieved high predictive accuracy due to advancements in data availability and increased computing power. In this paper, we present a caching algorithm called the user preference-aware content caching algorithm (UPACA). This algorithm is specifically designed for an edge content delivery platform where users can access content services provided by a remote content provider. UPACA operates in two steps. In the first step, the proposed collaborative filtering-based popularity prediction algorithm (CFPA) is used to predict future content popularities. CFPA utilizes a gated residual variational autoencoder collaborative filtering model to predict users' future preferences and calculate the future popularity of content. This algorithm considers the popularity of the content as well as the number and timing of content requests. Experimental results demonstrate that UPACA outperforms previous methods in terms of cache hit rates and user utilities. [ABSTRACT FROM AUTHOR]

A neural network model needs to be manually designed for ancient mural dynasty recognition, and this paper proposes an ancient mural dynasty recognition algorithm that is based on a neural architecture search (NAS). First, the structural edge information of mural images is extracted for use by the neural network model in recognizing mural missions. Second, an NAS algorithm that is based on contrast selection (CS) simplifies the architecture search to an incremental CS and then searches for the optimal network architecture on the mural dataset. Finally, the identified optimal network architecture is used for training and testing to complete the mural dynasty recognition task. The results show that the top accuracy of the proposed method on the mural dataset is 88.10%, the recall rate is 87.52%, and the precision rate is 87.69%. Each evaluation index used by the neural network model is superior to that of classical network models such as AlexNet and ResNet-50. Compared with NAS methods such as ASNG and MIGO, the accuracy of mural dynasty recognition is higher by an average of 4.27% when using the proposed method. The proposed method is verified on CIFAR-10, CIFAR-100, ImageNet16-120 and other datasets and achieves a good recognition accuracy in the NAS-bench-201 search space, which averages 93.26%, 70.73% and 45.34%, respectively, on the abovementioned datasets. [ABSTRACT FROM AUTHOR]

This paper aims to propose a table tennis target detection (TD) method based on deep learning (DL) and multi-scale feature fusion (MFF) to improve the detection accuracy of the ball in table tennis competition, optimize the training process of athletes, and improve the technical level. In this paper, DL technology is used to improve the accuracy of table tennis TD through MFF guidance. Initially, based on the FAST Region-based Convolutional Neural Network (FAST R-CNN), the TD is carried out in the table tennis match. Then, through the method of MFF guidance, different levels of feature information are fused, which improves the accuracy of TD. Through the experimental verification on the test set, it is found that the mean Average Precision (mAP) value of the target detection algorithm (TDA) proposed here reaches 87.3%, which is obviously superior to other TDAs and has higher robustness. The DL TDA combined with the proposed MFF can be applied to various detection fields and can help the application of TD in real life. [ABSTRACT FROM AUTHOR]

Multimedia data have increased dramatically today, making the distinction between desirable information and other types of information extremely important. Speech/music discrimination is a field of audio analytics that aims to detect and classify speech and music segments in an audio file. This paper proposes a novel feature extraction method called Long-Term Multi-band Frequency-Domain Mean-Crossing Rate (FDMCR). The proposed feature computes the average frequency-domain mean-crossing rate along the frequency axis for each of the perceptual Mel-scaled frequency bands of the signal power spectrum. In this paper, the class-separation capability of this feature is first measured by well-known divergence criteria such as Maximum Fisher Discriminant Ratio (MFDR), Bhattacharyya divergence, and Jeffreys/Symmetric Kullback–Leibler (SKL) divergence. The proposed feature is then applied to the speech/music discrimination (SMD) process on two well-known speech-music datasets—GTZAN and S &S (Scheirer and Slaney). The results obtained on the two datasets using conventional classifiers, including k-NN, GMM, and SVM, as well as deep learning-based classification methods, including CNN, LSTM, and BiLSTM, show that the proposed feature outperforms other features in speech/music discrimination. [ABSTRACT FROM AUTHOR]

During the manufacturing process of printed circuit boards (PCBs), quality defects can occur, which can affect the performance and reliability of PCBs. Existing deep learning-based PCB defect detection methods are difficult to simultaneously achieve the goals of high detection accuracy, fast detection speed, and small number of parameters. Therefore, this paper proposes a PCB defect detection algorithm based on CDI-YOLO. Firstly, the coordinate attention mechanism (CA) is introduced to improve the backbone and neck network of YOLOv7-tiny, enhance the feature extraction capability of the model, and thus improve the accuracy of model detection. Secondly, DSConv is used to replace part of the common convolution in YOLOv7-tiny to achieve lower computing costs and faster detection speed. Finally, Inner-CIoU is used as the bounding box regression loss function of CDI-YOLO to speed up the bounding box regression process. The experimental results show that the method achieves 98.3% mAP on the PCB defect dataset, the detection speed is 128 frames per second (FPS), the parameters is 5.8 M, and the giga floating-point operations per second (GFLOPs) is 12.6 G. Compared with the existing methods, the comprehensive performance of this method has advantages. [ABSTRACT FROM AUTHOR]

The problem that I present in this paper concerns the issue of ethical evaluation of algorithms, especially those used in social media and which create profiles of users of these media and new technologies that have recently emerged and are intended to change the functioning of technologies used in data management. Systems such as Overton, SambaNova or Snorkel were created to help engineers create data management models, but they are based on different assumptions than the previous approach in machine learning and deep learning. There is a need to analyze both deep learning algorithms and new technologies in database management in terms of their actions towards a person who leaves their digital footprints, on which these technologies work. Then, the possibilities of applying the existing deep learning technology and new Big Data systems in the economy will be shown. The opportunities offered by the systems mentioned above seem to be promising for many companies and—if implemented on a larger scale—they will affect the functioning of the free market. [ABSTRACT FROM AUTHOR]

Distributed learning, as the most popular solution for training large-scale data for deep learning, consists of multiple participants collaborating on data training tasks. However, the malicious behavior of some during the training process, like Byzantine participants who would interrupt or control the learning process, will trigger the crisis of data security. Although recent existing defense mechanisms use the variability of Byzantine node gradients to clear Byzantine values, it is still unable to identify and then clear the delicate disturbance/attack. To address this critical issue, we propose an algorithm named consensus aggregation in this paper. This algorithm allows computational nodes to use the information of verification nodes to verify the effectiveness of the gradient in the perturbation attack, reaching a consensus based on the effective verification of the gradient. Then the server node uses the gradient as the valid gradient for gradient aggregation calculation through the consensus reached by other computing nodes. On the MNIST and CIFAR10 datasets, when faced with Drift attacks, the proposed algorithm outperforms common existing aggregation algorithms (Krum, Trimmed Mean, Bulyan), with accuracies of 93.3%, 94.06% (MNIST dataset), and 48.66%, 51.55% (CIFAR10 dataset), respectively. This is an improvement of 3.0%, 3.8% (MNIST dataset), and 19.0%, 26.1% (CIFAR10 dataset) over the current state-of-the-art methods, and successfully defended against other attack methods. [ABSTRACT FROM AUTHOR]

Adaptive algorithms are widely used in deep learning because of their fast convergence. Among them, Adam is the most widely used algorithm. However, studies have shown that Adam's generalization ability is weak. AdaX is a variant of Adam, which introduces a novel second-order momentum, modifies the second-order moment of Adam, and has good generalization ability. However, these algorithms may fail to converge due to instability and extreme learning rates during training. In this paper, we propose a new adaptive and momental bound algorithm, called AdaXod, which characterizes of exponentially averaging the learning rate and is particularly useful for training deep neural networks. By setting an adaptively limited learning rate in the AdaX algorithm, the resultant AdaXod can effectively eliminate the problem of excessive learning rate in the later stage of neural networks training and thus results in stable training. We conduct extensive experiments on different datasets and verify the advantages of the AdaXod algorithm by comparing with other advanced adaptive optimization algorithms. AdaXod eliminates large learning rates during neural networks training and outperforms other optimizers, especially for some neural networks with complex structures, such as DenseNet. [ABSTRACT FROM AUTHOR]

The ridge beast is a beast placed on the ridge of the roof of ancient Chinese buildings, not only has a decorative function, and has a strict hierarchical meaning, the number and form of the ridge beast placed on different levels of buildings are strictly limited. The detection technology of ridge beast decorative parts has important application value in the fields of fine 3D reconstruction of ancient buildings, historical dating and cultural and tourism services. Aiming at the problem of poor detection performance of traditional detection algorithms due to high texture similarity and poor discrimination of ridge beast, this paper proposed an improved YOLOv3 based detection algorithm for ridge beast decorative pieces. In terms of basic network improvement, local features are aggregated to the deep separable convolution internal embedding summation layer, and point convolution is used to connect the channel information of original features and aggregated features, so as to expand the receptive field and learn more diverse features. The residual structure of the feature extraction network was constructed by using the convolution, and the extraction effect of the model on the fine-grained features of the ridge beast was optimized, so that the detection accuracy was improved. In the prediction head improvement of the model, the original linear structure was reconstructed, and the extrusion and excitation modules were introduced to model the channel relationship of multi-scale feature map, which suppressed the response of interference signals and made the feature more directivity. The parallel 1 × 1 and 3 × 3 convolution are used to construct a multi-size convolution structure, which enhances the semantic information extraction ability of the model and further improves the detection effect. Experiments were conducted on the constructed ridge-beast dataset, and the results showed that the mAP of the improved algorithm can reach 86.48%, which is 3.05% higher than YOLO-v3, and the model parameters are reduced by 70%, which has a better detection performance and can provide a reference for the automated detection of ancient building components. [ABSTRACT FROM AUTHOR]

Distributed learning, as the most popular solution for training large-scale data for deep learning, consists of multiple participants collaborating on data training tasks. However, the malicious behavior of some during the training process, like Byzantine participants who would interrupt or control the learning process, will trigger the crisis of data security. Although recent existing defense mechanisms use the variability of Byzantine node gradients to clear Byzantine values, it is still unable to identify and then clear the delicate disturbance/attack. To address this critical issue, we propose an algorithm named consensus aggregation in this paper. This algorithm allows computational nodes to use the information of verification nodes to verify the effectiveness of the gradient in the perturbation attack, reaching a consensus based on the effective verification of the gradient. Then the server node uses the gradient as the valid gradient for gradient aggregation calculation through the consensus reached by other computing nodes. On the MNIST and CIFAR10 datasets, when faced with Drift attacks, the proposed algorithm outperforms common existing aggregation algorithms (Krum, Trimmed Mean, Bulyan), with accuracies of 93.3%, 94.06% (MNIST dataset), and 48.66%, 51.55% (CIFAR10 dataset), respectively. This is an improvement of 3.0%, 3.8% (MNIST dataset), and 19.0%, 26.1% (CIFAR10 dataset) over the current state-of-the-art methods, and successfully defended against other attack methods. [ABSTRACT FROM AUTHOR]

Ultrasonic vocalizations (USVs) analysis represents a fundamental tool to study animal communication. It can be used to perform a behavioral investigation of mice for ethological studies and in the field of neuroscience and neuropharmacology. The USVs are usually recorded with a microphone sensitive to ultrasound frequencies and then processed by specific software, which help the operator to identify and characterize different families of calls. Recently, many automated systems have been proposed for automatically performing both the detection and the classification of the USVs. Of course, the USV segmentation represents the crucial step for the general framework, since the quality of the call processing strictly depends on how accurately the call itself has been previously detected. In this paper, we investigate the performance of three supervised deep learning methods for automated USV segmentation: an Auto-Encoder Neural Network (AE), a U-NET Neural Network (UNET) and a Recurrent Neural Network (RNN). The proposed models receive as input the spectrogram associated with the recorded audio track and return as output the regions in which the USV calls have been detected. To evaluate the performance of the models, we have built a dataset by recording several audio tracks and manually segmenting the corresponding USV spectrograms generated with the Avisoft software, producing in this way the ground-truth (GT) used for training. All three proposed architectures demonstrated precision and recall scores exceeding 90 % , with UNET and AE achieving values above 95 % , surpassing other state-of-the-art methods that were considered for comparison in this study. Additionally, the evaluation was extended to an external dataset, where UNET once again exhibited the highest performance. We suggest that our experimental results may represent a valuable benchmark for future works. [ABSTRACT FROM AUTHOR]

Video Snapshot compressive imaging (SCI) is a promising technique to capture high-speed videos, which transforms the imaging speed from the detector to mask modulating and only needs a single measurement to capture multiple frames. The algorithm to reconstruct high-speed frames from the measurement plays a vital role in SCI. In this paper, we consider the promising reconstruction algorithm framework, namely plug-and-play (PnP), which is flexible to the encoding process comparing with other deep learning networks. One drawback of existing PnP algorithms is that they use a pre-trained denoising network as a plugged prior while the training data of the network might be different from the task in real applications. Towards this end, in this work, we propose the online PnP algorithm which can adaptively update the network's parameters within the PnP iteration; this makes the denoising network more applicable to the desired data in the SCI reconstruction. Furthermore, for color video imaging, RGB frames need to be recovered from Bayer pattern or named demosaicing in the camera pipeline. To address this challenge, we design a two-stage reconstruction framework to optimize these two coupled ill-posed problems and introduce a deep demosaicing prior specifically for video demosaicing in SCI. Extensive results on both simulation and real datasets verify the superiority of our adaptive deep PnP algorithm. Code is available at https://github.com/xyvirtualgroup/AdaptivePnP_SCI. [ABSTRACT FROM AUTHOR]

Based on the problem of insufficient accuracy of the original tiny YOLOv3 algorithm for object detection in a lawn environment, an Optimized tiny YOLOv3 algorithm with less computation and higher accuracy is proposed. Three reasons affect the accuracy of the original tiny YOLOv3 algorithm for detecting objects in a lawn environment. First, the backbone of the original algorithm is composed of a stack of a single convolutional layer and a max-pooling layer, which results in insufficient ability to extract feature information of objects. An enhancement module is proposed to enhance the feature extraction capability of the shallow layers of the network. Second, the information of the shallow convolutional layers of the backbone is not fully used, which results in insufficient detection capability for small objects. Third, the deep part of the backbone uses a convolutional layer with an excessive number of channels, which results in a large amount of computation. A multi-resolution fusion module is proposed to enhance the information interaction capability of the deep and shallow layers of the network, and reduce the computation. To verify the accuracy of this Optimized tiny YOLOv3 algorithm, the algorithm was tested on the dataset containing trunk, spherical tree and person, and compared with the current research. The results show that the algorithm proposed in this paper improves the detection accuracy while reducing the calculation. [ABSTRACT FROM AUTHOR]

This paper aims to improve the quality and efficiency of talent cultivation in colleges and universities and effectively analyze and count the number of talents. The cultivation of computer professionals in higher vocational education is taken as the research object. First, a talent cultivation information system for colleges and universities is developed, using deep learning neural networks and data mining algorithms and introducing recurrent neural networks (RNN) and a fusion attention model. With a view to the needs of employment forms and employment data, a diversified talent demand cultivation program is formulated. Based on this, an efficient talent cultivation quality indicator system is established through analyzing talent cultivation data in higher vocational education. Finally, the effectiveness of the system is verified by the performance analysis. The results show an average prediction accuracy for the proposed Attention RNN Word2 algorithm of 80.61%, which is significantly higher than achieved by the latest research algorithm model, and the talent training has a classification accuracy of 82%. This model is not only feasible for categorizing computer professional personnel cultivation but can also accurately reflect problems arising in personnel cultivation, and the results are made obvious by visualization. To sum up, the system reported here can give an intuitive demonstration of the quality of computer professional personnel cultivation in higher vocational education. This research provides a theoretical basis for related research into the cultivation of computer professionals in higher vocational education for application in industry, universities, and research. [ABSTRACT FROM AUTHOR]

Due to the fact that coverless information hiding can effectively resist the detection of steganalysis tools, it has attracted more attention in the field of information hiding. At present, most coverless information hiding schemes select text and image as transmission carriers, while there are few studies on emerging popular media such as video, which has more abundant contents. Taking the natural video as the carrier is more secure and can avoid the attention of attackers. In this paper, we propose a coverless video steganography algorithm based on semantic segmentation. Specifically, to establish the mapping relationship between secret information and video files effectively, this paper introduces the deep learning based on semantic segmentation network to calculate the statistical histogram of semantic information. To quickly index the sender's secret message to the corresponding video frame, we build a three-digit index structure. The receiver can extract the valid video frame from the three-digit index information and restore the secret information. On the one hand, the neural network is trained through the original image and the noisy image in this scheme; therefore, it can not only effectively resist the interference of noises, but also accurately extract the robust deep features of the image. The frames of video generate the robust mapping to the secret information after the semantic information statistics. On the other hand, semantic segmentation belongs to pixel-level segmentation, which has high requirements for network parameters, so it is difficult for attackers to decrypt and recover secret information. Since this scheme does not modify the primitiveness of video data, it can effectively resist steganalysis tools. The experimental results and analysis show that the video coverless information hiding scheme has a large capacity and a certain resistance to noise attack. [ABSTRACT FROM AUTHOR]

With the rapid development of computer hardware in the past three decades, various classic algorithms such as neural computing and bionic optimization computing have been widely used in practical problems. This paper extended the new bionic algorithm-flock algorithm proposed in 2014 and obtained a multi-objective flock algorithm to solve the multi-objective problem. This study used aggregate functions to define social ranks, and simulated the foraging behavior of chickens in the process of searching for food in the objective space and found the balance between diversity and convergence when looking for the best Pareto solution. The algorithm took five types of bi-objective functions and four types of three-objective functions as objects and compared it with four more widely used algorithms in multi-objective problems. The results demonstrate that the MOCSO (multi-objective chicken swarm optimization) algorithm shows better results in the optimization of multi-objective problems. [ABSTRACT FROM AUTHOR]

In this paper, we consider the problem of sparse signal recovery using a learned dictionary in multiple measurement vectors (MMVs) case. Employing deep neural networks, we provide two new greedy algorithms to solve sparse MMV problems. In the first algorithm, we create a stacked vector of measurement matrix columns and a new measurement matrix, which can be assumed as the Kronecker product of the primary compressive sampling matrix and a unitary matrix. In order to reconstruct sparse vector corresponding to this new set of equations, a four-layer feed-forward neural network is applied. In the second algorithm, joint sparse structure of the sparse vectors is considered. Recurrent neural networks are employed to extract the joint sparsity structure. In addition, we utilize an over-complete dictionary obtained from an unsupervised learning procedure. Simulation results illustrate the benefit of using the proposed methods. Finally, the proposed algorithms are applied for pilot-based channel estimation in massive multiple-input multiple-output systems to improve the channel state information recovery performance. [ABSTRACT FROM AUTHOR]

This work focuses on document fragments association using deep metric learning methods. More precisely, we are interested in ancient papyri fragments that need to be reconstructed prior to their analysis by papyrologists. This is a challenging task to automatize using machine learning algorithms because labeled data is rare, often incomplete, imbalanced and of inconsistent conservation states. However, there is a real need for such software in the papyrology community as the process of reconstructing the papyri by hand is extremely time-consuming and tedious. In this paper, we explore ways in which papyrologists can obtain useful matching suggestion on new data using Deep Convolutional Siamese-Networks. We emphasize on low-to-no human intervention for annotating images. We show that the from-scratchself-supervised approach we propose is more effective than using knowledge transfer from a large dataset, the former achieving a top-1 accuracy score of 0.73 on a retrieval task involving 800 fragments. [ABSTRACT FROM AUTHOR]

Information leaks via side channels remain a challenging problem to guarantee confidentiality. Static analysis is a prevalent approach for detecting side channels. However, the side-channel analysis poses challenges to the static techniques since they arise from non-functional aspects of systems and require an analysis of multiple traces. In addition, the outcome of static analysis is usually restricted to binary answers. In practice, real-world applications may need to disclose some aspects of the confidential information to ensure desired functionality. Therefore, quantification techniques are necessary to evaluate the resulting threats. In this paper, we propose a dynamic analysis technique to detect and quantify side channels. Our novel approach is to split the problem into two tasks. First, we learn a timing model of the program as a neural network. While the program implements the functionality, the neural network models the non-functional property that does not exist in the syntax or semantics of programs. Second, we analyze the neural network to quantify information leaks. As demonstrated in our experiments, both of these tasks are feasible in practice—making the approach a significant improvement over state-of-the-art side channel detectors and quantifiers. Thus, our key technical contributions are (a) a binarized neural network architecture that enables side-channel discovery and (b) a novel MILP-based counting algorithm to estimate the side-channel strength. On a set of benchmarks, we show that neural network models the timing of programs with thousands of methods precisely. We also show that neural networks with thousands of neurons can be efficiently analyzed to quantify information leaks via timing side channels. [ABSTRACT FROM AUTHOR]

Since the particle swarm optimization (PSO) was proposed to overcome the inherent defects of PSO such as premature convergence and dependent on parameters settings, different PSO variants are devised to optimize different complex optimization problems; NOPSO is an excellent representative among them. This paper ensembles the three types of velocity update formulas proposed in NOPSO and presents a dual-drive opposition-based non-inertial PSO to improve the robustness of algorithm while ensuring the searching efficiency and solving accuracy of optimization procedure. Two main strategies are introduced in the new algorithm: (1) a dual-drive velocity update formula (DDVM) is proposed to control move of particles and (2) an elite differential evolutionary mutation strategy (EDEM) is devised to help particles escape from local optimum. The modified algorithm with two above strategies is proved to be competitive compared with some state-of-the-art OBL-based PSO including NOPSO and can be effectively applied to deep learning in IoTs in the foreseeable future. [ABSTRACT FROM AUTHOR]

Image registration is a fundamental task in image analysis in which the transform that moves the coordinate system of one image to another is calculated. Registration of multi-modal medical images has important implications for clinical diagnosis, treatment planning, and image-guided surgery as it provides the means of bringing together complimentary information obtained from different image modalities. However, since different image modalities have different properties due to their different acquisition methods, it remains a challenging task to find a fast and accurate match between multi-modal images. Furthermore, due to reasons such as ethical issues and need for human expert intervention, it is difficult to collect a large database of labelled multi-modal medical images. In addition, manual input is required to determine the fixed and moving images as input to registration algorithms. In this paper, we address these issues and introduce a registration framework that (1) creates synthetic data to augment existing datasets, (2) generates ground truth data to be used in the training and testing of algorithms, (3) registers (using a combination of deep learning and conventional machine learning methods) multi-modal images in an accurate and fast manner, and (4) automatically classifies the image modality so that the process of registration can be fully automated. We validate the performance of the proposed framework on CT and MRI images of the head obtained from a publicly available registration database. [ABSTRACT FROM AUTHOR]

Three-dimensional object reconstruction from multi-view images is an important topic in computer vision, which has attracted enormous attention during the past decades. With the further study in deep learning, remarkable progress of three-dimensional object reconstruct has been obtained in recent years. In this paper, we proposed three-dimensional rapid registration and reconstruction of multi-view rigid objects based on end-to-end deep surface model in the field of three-dimensional object reconstruction. Firstly, we introduce a matching algorithm called local stereo matching algorithm based on improved census transform and multi-scale spatial, aiming to improve the matching results for those regions. In cost aggregation step, guided map filtering algorithm with excellent gradient preserving property is introduced into Gaussian pyramid structure and regularization is added to strengthen cost volume consistency. Secondly, the improved inception RESNET module is added to improve the feature extraction ability of the network, and multiple features are extracted by using multiple network structures, and finally multiple features are sequentially input into the VRNN module to enhance the reconstruction effect of multi-view images. The experimental results show that our proposed method can not only achieve better reconstruction results, but also reconstruct more details and spend less time in training. [ABSTRACT FROM AUTHOR]

Removing motion blur has been an important issue in computer vision literature. Motion blur is caused by the relative motion between the camera and the photographed object. However, in recent years, some achievements have been made in the research of image deblurring by using deep learning algorithms. In this paper, an enhanced adversarial network model is proposed. The proposed model can use the weight of feature channel to generate sharp image and eliminate draughtboard artefacts. In addition, the mixed loss function enables the network to output high-quality image. The proposed approach is tested using GOPRO datasets and Lai datasets. In the GOPRO datasets, the peak signal-to-noise ratio of the proposed approach is up to 28.674, and DeblurGAN is 27.454. And the structural similarity measure can be achieved up to 0.969, and DeblurGAN is 0.939. Furthermore, the images were obtained from China's Chang'e 3 Lander to test the new algorithm. Due to the elimination of the chessboard effect, the deblurred image has a better visual appearance. The proposed method achieved higher performance and efficiency in qualitative and quantitative aspects using the benchmark dataset experiments. The results also provided various insights into the design and development of the camera pointing system, which was mounted on the Lander for capturing images of the moon and rover for Chang'e space mission. [ABSTRACT FROM AUTHOR]

Training deep learning model is a time-consuming job since it usually uses a large amount of data. To reduce the training time, most practitioners train the models in a GPU cluster environment in a distributed way. The synchronous stochastic gradient descent, which is one of the widely used distributed training algorithms, has fast convergence rate with the use of multiple GPU workers; but its speed is tied to the slowest worker, i.e., straggler. In a heterogeneous environment, a static straggler, which has not been mainly focused before, has more impact on the performance than randomly occurring straggler. However, most existing studies for straggler mitigation usually consider a homogeneous environment, so their approaches are limited in practice. In this paper, we scrutinize the straggler problem under heterogeneous environment and define static and dynamic straggler from empirical results. Based on this, we propose a novel approach called batch orchestration algorithm (BOA) for straggler mitigation. It adaptively balances the amount of mini-batch data according to the speed of workers. Therefore, BOA can mitigate both static and dynamic straggler in a modern GPU cluster. BOA uses a Min–Max Integer programming to find the optimal mini-batch size, with the hardware-agnostic performance models. For verification, several experiments are conducted on a cluster having up to six GPUs with three types: GTX 1080, GTX 1060 and Quadro M2000. The results show BOA mitigates both types of stragglers and accelerates the training speed with synchronous SGD compared to other straggler mitigation method. [ABSTRACT FROM AUTHOR]

Deep neural network (DNN) architecture based models have high expressive power and learning capacity. However, they are essentially a black box method since it is not easy to mathematically formulate the functions that are learned within its many layers of representation. Realizing this, many researchers have started to design methods to exploit the drawbacks of deep learning based algorithms questioning their robustness and exposing their singularities. In this paper, we attempt to unravel three aspects related to the robustness of DNNs for face recognition: (i) assessing the impact of deep architectures for face recognition in terms of vulnerabilities to attacks, (ii) detecting the singularities by characterizing abnormal filter response behavior in the hidden layers of deep networks; and (iii) making corrections to the processing pipeline to alleviate the problem. Our experimental evaluation using multiple open-source DNN-based face recognition networks, and three publicly available face databases demonstrates that the performance of deep learning based face recognition algorithms can suffer greatly in the presence of such distortions. We also evaluate the proposed approaches on four existing quasi-imperceptible distortions: DeepFool, Universal adversarial perturbations, l 2 , and Elastic-Net (EAD). The proposed method is able to detect both types of attacks with very high accuracy by suitably designing a classifier using the response of the hidden layers in the network. Finally, we present effective countermeasures to mitigate the impact of adversarial attacks and improve the overall robustness of DNN-based face recognition. [ABSTRACT FROM AUTHOR]

Join is an essential tool for data analysis which collected from different data sources. MapReduce has emerged as a prominent programming model for processing of massive data. However, traditional join algorithms based on MapReduce are not efficient when handling skewed data. The presence of data skew in input data leads to considerable load imbalance and performance degradation. This paper proposes a new skew-insensitive method, called fine-grained partitioning for skew data (FGSD) which can improve the load balancing for reduce tasks. The proposed method considers the properties of both input and output data through a proposed stream sampling algorithm. FGSD introduces a new approach for distribution of input data which leads to efficiently handling redistribution and join product skew. The experimental results confirm that our solution can not only achieve higher balancing performance, but also reduce the execution time of a job with varying degrees of the data skew. Furthermore, FGSD does not require any modification to the MapReduce environment and is applicable to complex join. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNNs) is one of the most popular machine learning methods and is widely used in many modern applications. The training process of DNNs is a time-consuming process. Accelerating the training of DNNs has been the focus of many research works. In this paper, we speed up the training of DNNs applied for automatic speech recognition and the target architecture is heterogeneous (CPU + MIC). We apply asynchronous methods for I/O and communication operations and propose an adaptive load balancing method. Besides, we also employ a momentum idea to speed up the convergence of the gradient descent algorithm. Experimental results show that our optimized algorithm is able to acquire a 20-fold speedup on a CPU + MIC platform compared with the original sequential algorithm on a single-core CPU. [ABSTRACT FROM AUTHOR]

Artificial intelligence has been widely studied on solving intelligent surveillance analysis and security problems in recent years. Although many multimedia security approaches have been proposed by using deep learning network model, there are still some challenges on their performances which deserve in-depth research. On the one hand, high computational complexity of current deep learning methods makes it hard to be applied to real-time scenario. On the other hand, it is difficult to obtain the specific features of a video by fine-tuning the network online with the object state of the first frame, which fails to capture rich appearance variations of the object. To solve above two issues, in this paper, an effective object tracking method with learning attention is proposed to achieve the object localization and reduce the training time in adversarial learning framework. First, a prediction network is designed to track the object in video sequences. The object positions of the first ten frames are employed to fine-tune prediction network, which can fully mine a specific features of an object. Second, the prediction network is integrated into the generative adversarial network framework, which randomly generates masks to capture object appearance variations via adaptively dropout input features. Third, we present a spatial attention mechanism to improve the tracking performance. The proposed network can identify the mask that maintains the most robust features of the objects over a long temporal span. Extensive experiments on two large-scale benchmarks demonstrate that the proposed algorithm performs favorably against state-of-the-art methods. [ABSTRACT FROM AUTHOR]

To automatically detect dynamic EEG signals to reduce the time cost of epilepsy diagnosis. In the signal recognition of electroencephalogram (EEG) of epilepsy, traditional machine learning and statistical methods require manual feature labeling engineering in order to show excellent results on a single data set. And the artificially selected features may carry a bias, and cannot guarantee the validity and expansibility in real-world data. In practical applications, deep learning methods can release people from feature engineering to a certain extent. As long as the focus is on the expansion of data quality and quantity, the algorithm model can learn automatically to get better improvements. In addition, the deep learning method can also extract many features that are difficult for humans to perceive, thereby making the algorithm more robust. Based on the design idea of ResNeXt deep neural network, this paper designs a Time-ResNeXt network structure suitable for time series EEG epilepsy detection to identify EEG signals. The accuracy rate of Time-ResNeXt in the detection of EEG epilepsy can reach 91.50%. The Time-ResNeXt network structure produces extremely advanced performance on the benchmark dataset (Berne-Barcelona dataset) and has great potential for improving clinical practice. [ABSTRACT FROM AUTHOR]