Search

Electroencephalogram (EEG)-based emotion recognition has become a hot research field, with the most attention given to decoding three basic types of emotional states (i.e., positive, negative, and neutral) by EEG. Traditional EEG emotion recognition is a single feature input mode that cannot cover multiple feature information. For brain function networks, the feature extraction process is very cumbersome. There will also be individual differences in the characteristics of different subjects. Therefore, the theory of graph convolution and brain function connection is introduced into this research, and the multi-domain fusion features input deep graph convolution neural network (MdGCNN) is proposed in this paper. Pearson correlation is used to determine the adjacency matrix. The Sortpooling layer is employed as a bridge between the graph convolution neural layer and the normal neural network layer and sorts the node features in a consistent order. Based on analyzing the characteristics of a single electrode, the brain topology structure features are automatically extracted. Taking MdGCNN as the basic model and considering the method of minimizing the feature distance between the source and the target domain, we propose a transfer learning (TL) emotion recognition model for cross-subject called MdGCNN-TL. Meanwhile, MdGCNN-TL is extended to traverse the target domain of a single subject in a two-to-one domain form. According to the idea of principal component analysis, the transfer model with a high recognition effect is determined with the degree of subject correlation (SC), and MdGCNN-TL is upgraded to MdGCNN-TL-SC. Experimental analysis on the SEED dataset is performed to evaluate the proposed models. Further validation of the model is implemented on the DEAP dataset. The results show that the proposed model has achieved better performance in EEG emotion recognition. [ABSTRACT FROM AUTHOR]

Rapid advances in deep learning models have made it easier for public and crackers to generate hyper-realistic deepfake videos in which faces are swapped. Such deepfake videos may constitute a significant threat to the world if they are misused to blackmail public figures and to deceive systems of face recognition. As a result, distinguishing these fake videos from real ones has become fundamental. This paper introduces a new deepfake video detection method. You Only Look Once (YOLO) face detector is used to detect faces from video frames. A proposed hybrid method based on proposing two different feature extraction methods is applied to these faces. The first feature extraction method, a proposed Convolution Neural Network (CNN), is based on the Histogram of Oriented Gradient (HOG) method. The second one is an ameliorated XceptionNet CNN. The two extracted sets of features are merged together and fed as input to a sequence of Gated Recurrent Units (GRUs) to extract the spatial and temporal features and then individuate the authenticity of videos. The proposed method is trained on the CelebDF-FaceForencics++ (c23) dataset and evaluated on the CelebDF test set. The experimental results and analysis confirm the superiority of the suggested method over the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Identifying distortions in images and videos is important and useful in various visual applications, such as image quality enhancement and assessment techniques. Instead of applying them blindly, these techniques can be applied or adjusted depending on the type of distortion identified. In this paper, we propose a deep multi-task learning (MTL) model for identifying the types of distortion in both images and videos, considering both single and multiple distortions. The proposed MTL model is composed of one convolutional neural network (CNN) shared between all tasks and N parallel classifiers, where each classifier is dedicated to identify a type of distortion. The proposed architecture also allows to adjust the number of tasks according to the number of distortion types considered, making the solution scalable. The proposed method has been evaluated on natural scene images and laparoscopic videos databases, each presenting a rich set of distortions. The experimental results demonstrate that our model achieves the best performance among the state-of-the-art methods for both single and multiple distortions (Code is available at: https://github.com/zoubidaameur/Deep-Multi-Task-Learning-for-Image-Video-Distortions-Identification). [ABSTRACT FROM AUTHOR]

Emotion is an instinctive or intuitive feeling as distinguished from reasoning or knowledge. It varies over time, since it is a natural instinctive state of mind deriving from one's circumstances, mood, or relationships with others. Since emotions vary over time, it is important to understand and analyze them appropriately. Existing works have mostly focused well on recognizing basic emotions from human faces. However, the emotion recognition from cartoon images has not been extensively covered. Therefore, in this paper, we present an integrated Deep Neural Network (DNN) approach that deals with recognizing emotions from cartoon images. Since state-of-works do not have large amount of data, we collected a dataset of size 8 K from two cartoon characters: 'Tom' & 'Jerry' with four different emotions, namely happy, sad, angry, and surprise. The proposed integrated DNN approach, trained on a large dataset consisting of animations for both the characters (Tom and Jerry), correctly identifies the character, segments their face masks, and recognizes the consequent emotions with an accuracy score of 0.96. The approach utilizes Mask R-CNN for character detection and state-of-the-art deep learning models, namely ResNet-50, MobileNetV2, InceptionV3, and VGG 16 for emotion classification. In our study, to classify emotions, VGG 16 outperforms others with an accuracy of 96% and F1 score of 0.85. The proposed integrated DNN outperforms the state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

Despite the enormous performance of deep neural networks (DNNs), recent studies have shown their vulnerability to adversarial examples (AEs), i.e., carefully perturbed inputs designed to fool the targeted DNN. Currently, the literature is rich with many effective attacks to craft such AEs. Meanwhile, many defense strategies have been developed to mitigate this vulnerability. However, these latter showed their effectiveness against specific attacks and does not generalize well to different attacks. In this paper, we propose a framework for defending DNN classifier against adversarial samples. The proposed method is based on a two-stage framework involving a separate detector and a denoising block. The detector aims to detect AEs by characterizing them through the use of natural scene statistic (NSS), where we demonstrate that these statistical features are altered by the presence of adversarial perturbations. The denoiser is based on block matching 3D (BM3D) filter fed by an optimum threshold value estimated by a convolutional neural network (CNN) to project back the samples detected as AEs into their data manifold. We conducted a complete evaluation on three standard datasets, namely MNIST, CIFAR-10 and Tiny-ImageNet. The experimental results show that the proposed defense method outperforms the state-of-the-art defense techniques by improving the robustness against a set of attacks under black-box, gray-box and white-box settings. The source code is available at: https://github.com/kherchouche-anouar/2DAE. [ABSTRACT FROM AUTHOR]

Social media are the main contributors to spreading fake images. Fake images are manipulated images altered through software or by other means to change the information they convey. Fake images propagated over microblogging platforms generate misrepresentation and stimulate polarization in the people. Detection of fake images shared over social platforms is extremely critical to mitigating its spread. Fake images are often associated with textual data. Hence, a multi-modal framework is employed utilizing visual and textual feature learning. However, few multi-modal frameworks are already proposed; they are further dependent on additional tasks to learn the correlation between modalities. In this paper, an efficient multi-modal approach is proposed, which detects fake images of microblogging platforms. No further additional subcomponents are required. The proposed framework utilizes explicit convolution neural network model EfficientNetB0 for images and sentence transformer for text analysis. The feature embedding from visual and text is passed through dense layers and later fused to predict fake images. To validate the effectiveness, the proposed model is tested upon a publicly available microblogging dataset, MediaEval (Twitter) and Weibo, where the accuracy prediction of 85.3% and 81.2% is observed, respectively. The model is also verified against the newly created latest Twitter dataset containing images based on India's significant events in 2020. The experimental results illustrate that the proposed model performs better than other state-of-art multi-modal frameworks. [ABSTRACT FROM AUTHOR]

Age estimation from face images can be profitably employed in several applications, ranging from digital signage to social robotics, from business intelligence to access control. Only in recent years, the advent of deep learning allowed for the design of extremely accurate methods based on convolutional neural networks (CNNs) that achieve a remarkable performance in various face analysis tasks. However, these networks are not always applicable in real scenarios, due to both time and resource constraints that the most accurate approaches often do not meet. Moreover, in case of age estimation, there is the lack of a large and reliably annotated dataset for training deep neural networks. Within this context, we propose in this paper an effective training procedure of CNNs for age estimation based on knowledge distillation, able to allow smaller and simpler "student" models to be trained to match the predictions of a larger "teacher" model. We experimentally show that such student models are able to almost reach the performance of the teacher, obtaining high accuracy over the LFW+, LAP 2016 and Adience datasets, but being up to 15 times faster. Furthermore, we evaluate the performance of the student models in the presence of image corruptions, and we demonstrate that some of them are even more resilient to these corruptions than the teacher model. [ABSTRACT FROM AUTHOR]

Sea fog is a dangerous weather phenomenon that seriously affects maritime traffic and other operations at sea. The conventional sea fog detection methods are not only difficult to make full advantage of the multispectral information of cloud images, but also deficient in the exploration of deep-level semantic information, leading to poor detection results. In this paper, we proposed a multimodal self-supervised convolutional neural network incorporating intra-modal band attention mechanism (MSCNN-IBAM) based on multispectral images of Himawari-8. MSCNN-IBAM uses independent branches to extract features from different modality cloud images and characterize the importance of each band through attention mechanisms. Simultaneously, multimodal self-supervised learning and supervised learning are effectively combined to optimize the model by constructing a two-tuple trainset. Experimental results show the accuracy, precision, recall, and F1 score of the proposed method as 97.72%, 95.84%, 96.54%, and 96.08%, respectively, which have the competitive performance and acceptable computational efficiency. And the additional analysis of sea fog cases shows that the proposed method is not only effective in identifying sea fog, but also has the ability to locate sea fog regions. [ABSTRACT FROM AUTHOR]

Person re-identification (ReID) aims to match the specific pedestrians in the public environment with cross-domain cameras. Posture change, occlusion, and viewpoint change complicate ReID. Representation learning and metric learning have been focused on in previous researches as a key to solving the problem of ReID task by adaptive learning emphasized character feature representation. This paper proposed a Gaussian-based probability fusion mechanism and an improved Taylor Angular Margin softmax loss function, that the methods can be joined to learn more discriminative features and then better embedding space with categories weights. Different from the existing network attention intervention methods that by guided measurement quality fuse with prediction, the probability fusion structure can generate fine-grained recognition features via attention learning, and then send these features into the proposed Taylor angular margin loss. This proposed loss can push the inter-class margin and mining the intra-class variance for more effective performance. Massive experiments on Market-1501, DukeMTMC-reID, and MSMT17 show predominant performance promotion. It achieves 90.5% mAP and 96.4% Rank-1 on Market-1501, 80.5% mAP and 89.8% Rank-1 on DukeMTMC-reID 59.7% mAP and 82.5% Rank-1 on MSMT17. Moreover, excellent improvements are shown on occlusion datasets, 55.6% mAP and 65.1% Rank-1 on Occluded-DukeMTMC, 74.8% mAP and 80.5% Rank-1 on Occluded-ReID. [ABSTRACT FROM AUTHOR]

Adversarial examples can attack multiple unknown convolutional neural networks (CNNs) due to adversarial transferability, which reveals the vulnerability of CNNs and facilitates the development of adversarial attacks. However, most of the existing adversarial attack methods possess a limited transferability on vision transformers (ViTs). In this paper, we propose a partial blocks search attack (PBSA) method to generate adversarial examples on ViTs, which significantly enhance transferability. Instead of directly employing the same strategy for all encoder blocks on ViTs, we divide encoder blocks into two categories by introducing the block weight score and exploit distinct strategies to process them. In addition, we optimize the generation of perturbations by regularizing the self-attention feature maps and creating an ensemble of partial blocks. Finally, perturbations are adjusted by an adaptive weight to disturb the most effective pixels of original images. Extensive experiments on the ImageNet dataset are conducted to demonstrate the validity and effectiveness of the proposed PBSA. The experimental results reveal the superiority of the proposed PBSA to state-of-the-art attack methods on both ViTs and CNNs. Furthermore, PBSA can be flexibly combined with existing methods, which significantly enhances the transferability of adversarial examples. [ABSTRACT FROM AUTHOR]

The object tracking field continues to evolve as an important application of computer vision. Real-time performance is typically required in most applications of object tracking. The recent introduction of Convolutional Neural network (CNN) techniques to the object tracking field enabled the attainment of significant performance gains. However, the heavy computational load required for CNNs conflicts with the real-time requirements required for object tracking. In this paper, we address these computational limitations on the algorithm-side and the circuit-side. On the algorithm side, we adopt interpolation schemes which can significantly reduce the processing time and the memory storage requirements. We also evaluate the approximation of the hardware-expensive computations to attain an efficient hardware design. Moreover, we modify the online-training scheme in order to achieve a constant processing time across all video frames. On the circuit side, we developed a hardware accelerator of the online training stage. We avoid transposed reading from the external memory to speed-up the data movement with no performance degradation. Our proposed hardware accelerator achieves 44 frames-per-second in training the fully connected layers. [ABSTRACT FROM AUTHOR]

The growing usage of digital microphones has generated an increased interest in the topic of Acoustic Anomaly Detection (AAD). Indeed, there are several real-world AAD application domains, including working machines and in-vehicle intelligence (the main target of this research project). This paper introduces three deep AutoEncoders (AE) for unsupervised AAD tasks, namely a Dense AE, a Convolutional Neural Network (CNN) AE and Long Short-Term Memory Autoencoder (LSTM) AE. To tune the deep learning architectures, development data were adopted from public domain audio datasets related with working machines. A large set of computational experiments was held, showing that the three proposed deep autoencoders, when combined with a melspectrogram sound preprocessing, are quite competitive and outperform a recently proposed AE baseline. Next, on a second experimental stage, aiming to address the final in-vehicle passenger safety goal, the three AEs were adapted to learn from in-vehicle normal audio, assuming three realistic scenarios that were generated by a synthetic audio mixture tool. In general, a high quality AAD discrimination was obtained: working machine data – 72% to 91%; and in-vehicle audio – 78% to 81%. In conjunction with an automotive company, an in-vehicle AAD intelligent system prototype was further developed, aiming to test a selected model (LSTM AE) during a pilot demonstration event that targeted the cough anomaly. Interesting results were obtained, with the AAD system presenting a high cough classification accuracy (e.g., 100% for front seat locations). [ABSTRACT FROM AUTHOR]

In the previous works on heat stress detection, various state-of-the-art machine learning techniques had been used to detect stress patterns from electroencephalographic (EEG) signals. Since the handcrafted feature engineering-based approaches pose certain limitations and are sensitive to transform the nonlinearities of EEG, deep learning techniques have drawn attention in the domain of EEG-based applications. Moreover, existing approaches for stress detection consider the whole frequency band (delta to gamma) which conceals the redundant and lossy information that increased the false detection rate. Also, these approaches were implemented only to identify the stress in binary classes and confined their application to identifying the level of stress. Therefore, a multi-scale channel-wise convolution-based multi-level heat stress assessment has been proposed in this paper. The novel contributions of this work are (1) designing a multi-scale convolutional neural network (MS-CNN) for extracting precise information from individual frequency bands of EEG, (2) use of sparse connectivity to reduce the redundant information and increase the performance with less trainable parameters, and (3) multi-level heat stress assessment for precise identification of the severity of stress. The proposed approaches were evaluated on pre-recorded data of 10 rats in a simulated laboratory environment. The high accuracy of approximately 96–97% and 90–92% has been achieved for binary and multi-level stress detection. There is approximately a 6 to 50% reduction in the trainable parameters with a 2% improvement of accuracy achieved on adopting channel-wise convolution over MS-CNN and deep convolutional neural network (DCNN). This demonstrates the effectiveness of the adopted multiscale feature extraction and sparse connectivity to improve the performance with a less complex model. [ABSTRACT FROM AUTHOR]

In this paper, we introduce our unique dataset of fluorescence lifetime imaging endo/microscopy (FLIM), containing over 100,000 different FLIM images collected from 18 pairs of cancer/non-cancer human lung tissues of 18 patients by our custom fibre-based FLIM system. The aim of providing this dataset is that more researchers from relevant fields can push forward this particular area of research. Afterwards, we describe the best practice of image post-processing suitable per the dataset. In addition, we propose a novel hierarchically aggregated multi-scale architecture to improve the binary classification performance of classic CNNs. The proposed model integrates the advantages of multi-scale feature extraction at different levels, where layer-wise global information is aggregated with branch-wise local information. We integrate the proposal, namely ResNetZ, into ResNet, and appraise it on the FLIM dataset. Since ResNetZ can be configured with a shortcut connection and the aggregations by Addition or Concatenation, we first evaluate the impact of different configurations on the performance. We thoroughly examine various ResNetZ variants to demonstrate the superiority. We also compare our model with a feature-level multi-scale model to illustrate the advantages and disadvantages of multi-scale architectures at different levels. [ABSTRACT FROM AUTHOR]

Diabetic Retinopathy (DR), one of the most progressive sight-threatening diseases caused by the long-term diabetic condition, can lead to vision impairment and blindness later. Early diagnosis and timely treatment help control and avert DR from its progression. However, manual grading is exceptionally challenging and arduous due to the complex anatomical features in the retina. Therefore, developing an automated diagnostic method for screening DR is obligatory. This paper proposes a deep learning-enabled optimized feature selection approach to classify the stage of DR severity in the fundus image. At first, a pre-processing phase eradicates the noise and improves the contrast in the retinal fundus image. Then, blood vessel segmentation is performed using the Coherence Enhancing Energy Based Regularized Level Set Evolution method in the green channel fundus image. Subsequently, the optic disk is segmented with Canny Anisotropic Diffusion filter and morphological transformations. Next, the candidate lesion region is detected using an Attention-based Fusion Network (AFU-Net). Then, shape and texture features are extracted, and then, the optimal subset of features is selected using the Improved Harris Hawk Optimization algorithm. Finally, a deep Convolutional Neural Network classifies the DR stages, and the model weight is updated using the same algorithm. The proposed method achieved superior performance in two benchmark public datasets compared with the existing state-of-the-art methods using F1-score, accuracy, sensitivity, and specificity measures. [ABSTRACT FROM AUTHOR]

Object tracking has achieved impressive performance in computer vision. However, there are many challenges due to complex scenarios in reality. The mainstream trackers mostly locate the object in form of two branches, which limits the ability of trackers to fully mine similarity between template and search region. In this paper, we propose a multi-branch and multi-scale perception object tracking framework based on Siamese Convolutional Neural Networks (MultiBSP), in which the multi-branch tracking framework is established based on the idea of relation mining, and a tower-structured relation network is designed for each branch to learn the non-linear relation function between template and search region. By branch combination, multiple branches can verify their predictions with each other, which is beneficial to robust tracking. Besides, in order to sense the scale and aspect ratio of object in advance, a multi-scale perception module is designed by utilizing the dilated convolutions in five scales, which contributes to the ability of tracker to deal with scale variation. In addition, we propose an information enhancement module that focuses on important features and suppresses unnecessary ones along spatial and channel dimensions. Extensive experiments on six visual tracking benchmarks including OTB100, VOT2018, VOT2019, UAV123, GOT-10k, and LaSOT demonstrate that our MultiBSP can achieve robust tracking and have state-of-the-art performance. Finally, ablation experiments verify the effectiveness of each module and the tracking stability is proved by qualitative and quantitative analyses. [ABSTRACT FROM AUTHOR]

Financial forecasting has always been an intriguing research area in the field of finance. The widely accepted approach to forecast financial data is to perform predictions using time series data. In time series analysis, sampling the financial data with a predefined frequency (e.g. hourly, daily) leads to an uneven and discontinued data flow. Directional Change is a newly proposed approach that replaces physical time within the financial data and establishes an event-driven framework. With the emergence of the machine and deep learning-based methods, researchers have utilised them in financial time series. These techniques have shown to outperform conventional approaches. This paper aims to employ the CNN-LSTM model to investigate its predictive competence within the Directional Change (DC) framework to predict DC event prices. To obtain this objective, we first create the tick bars/candles of the GBPUSD, EURUSD, USDCHF, and USDCAD tick prices from January to August 2019. Then, the DC-based summaries of the selected tick bar/candle for each currency pair will be generated and fed to the CNN-LSTM model. The CNN-LSTM network architecture incorporates the robustness of Convolutional Neural Network (CNN) in feature extraction and Long Short-Term Memory (LSTM) in predicting sequential data. The results suggest that the performance of the CNN-LSTM model improves significantly within the DC framework. [ABSTRACT FROM AUTHOR]

This paper set out an automatic multicategory damage detection technique using convolutional neural networks (CNN) models based on image classification and features' extraction, to detect damages of historic structures such as: erosion, material loss, color change of the stone, and sabotage issues. The city of "Al-Salt" in Jordan was selected for the case study in this research. The best model showed an average damage detection accuracy of 95%. It was demonstrated that the proposed CNN model was significantly powerful, effective and reliable for damage detection of historic masonry buildings using features' extraction based on imaging, and it contributed to the management and safety of historic heritage and preservation. [ABSTRACT FROM AUTHOR]

Breast cancer is the second leading cause of death in women; therefore, effective early detection of this cancer can reduce its mortality rate. Breast cancer detection and classification in the early phases of development may allow for optimal therapy. Convolutional neural networks (CNNs) have enhanced tumor detection and classification efficiency in medical imaging compared to traditional approaches. This paper proposes a novel classification model for breast cancer diagnosis based on a hybridized CNN and an improved optimization algorithm, along with transfer learning, to help radiologists detect abnormalities efficiently. The marine predators algorithm (MPA) is the optimization algorithm we used, and we improve it using the opposition-based learning strategy to cope with the implied weaknesses of the original MPA. The improved marine predators algorithm (IMPA) is used to find the best values for the hyperparameters of the CNN architecture. The proposed method uses a pretrained CNN model called ResNet50 (residual network). This model is hybridized with the IMPA algorithm, resulting in an architecture called IMPA-ResNet50. Our evaluation is performed on two mammographic datasets, the mammographic image analysis society (MIAS) and curated breast imaging subset of DDSM (CBIS-DDSM) datasets. The proposed model was compared with other state-of-the-art approaches. The obtained results showed that the proposed model outperforms the compared state-of-the-art approaches, which are beneficial to classification performance, achieving 98.32% accuracy, 98.56% sensitivity, and 98.68% specificity on the CBIS-DDSM dataset and 98.88% accuracy, 97.61% sensitivity, and 98.40% specificity on the MIAS dataset. To evaluate the performance of IMPA in finding the optimal values for the hyperparameters of ResNet50 architecture, it compared to four other optimization algorithms including gravitational search algorithm (GSA), Harris hawks optimization (HHO), whale optimization algorithm (WOA), and the original MPA algorithm. The counterparts algorithms are also hybrid with the ResNet50 architecture produce models named GSA-ResNet50, HHO-ResNet50, WOA-ResNet50, and MPA-ResNet50, respectively. The results indicated that the proposed IMPA-ResNet50 is achieved a better performance than other counterparts. [ABSTRACT FROM AUTHOR]

Similarity searches on heterogeneous information networks (HINs) have attracted wide attention from both industrial and academic areas in recent years; for example, they have been used for friend detection in social networks and collaborator recommendation in coauthor networks. The structural information on the HIN can be captured by multiple metapaths, and people usually utilize metapaths to design methods for similarity search. The rich semantics in HINs are not only structural information but also content stored in nodes. However, the content similarity of nodes was usually not valued in the existing methods. Although some researchers have recently considered both types of information in machine learning-based methods for similarity search, they have used structure and content information separately. To address this issue by balancing the influence of structure and content information flexibly in the process of searching, we propose a double channel convolutional neural network model for top-k similarity search, which uses path instances as model inputs and generates structure and content embeddings for nodes based on different metapaths. We design an attention mechanism to enhance the differences in metapaths for each node. Another attention mechanism is used to combine the content and structure information of nodes. Finally, an importance evaluation function is designed to improve the accuracy and make the model more explainable. The experimental results show that our search algorithm can effectively support top-k similarity search in HINs and achieve higher performance than existing approaches. [ABSTRACT FROM AUTHOR]

Image dehazing is a challenging ill-posed problem in the field of computer vision. Existing learning-methods usually use a single convolutional neural network (CNN) model to solve it, which lacks details recovery mechanism and leads to poor performance. In this paper, we propose an end-to-end Dual-cascade Network for image dehazing, which obtains the haze-free image in a coarse-to-fine manner. Specifically, the overall model consists of two sub-networks: Net-U and Net-D. The Net-U employs an encoder–decoder architecture to restore a coarse dehazing result, which leverages residual channel attention block for distilling hierarchical features, and transmits the contextual information into next stage. To preserve the spatial details of latent image, our Net-D adopts a constant-size CNN structure, and captures the texture-rich features by utilizing residual multi-scale spatial block. Moreover, we apply an effective selective fusion module to integrate these derived features from Net-U and Net-D. Experimental comparisons show that our method obtains comparable or even better results than existing state-of-the-art methods in terms of quantitative evaluation and visual performance. The code will be made publicly available on GitHub. [ABSTRACT FROM AUTHOR]

Pedestrian Attribute Recognition (PAR) can provide valuable clues for several innovative surveillance applications. It is also a difficult task because inference of the multiple attributes at a far distance is challenging in real complex scenarios. Most existing methods improve the PAR with visual attention mechanisms or body-part detection modules, which increase the complexity of networks and require manual annotations of the human body. Also, uneven data distribution, leading to a decline in recall values, is still underestimated. This paper presents a novel multi-label optimization algorithm to mitigate these issues, named Multi-label Contrastive Focal Loss (MCFL). Specifically, we first propose a multi-label focal loss to emphasize the error-prone and minority attributes with a separated re-weighting scheme. And then, we introduce a multi-label contrastive learning strategy based on the multi-label divergences to help the deep network to distinguish the hard fine-grained attributes. We conduct extensive experiments on seven PAR benchmarks, and results indicate that the proposed MCFL with the native ResNet-50 backbone surpasses the state-of-the-art comparison methods in mean accuracy and recall. [ABSTRACT FROM AUTHOR]

Artificial neural network (ANN) is one of the technologies used for emerging real-world problems. Activation functions (AF) are used in deep learning architectures to make decisions in the hidden and output layers. An AF influences the dynamics of training and performance of an ANN. This paper proposes a novel AF called E-Tanh by extending the Tanh AF for ANN models. When used in three shallow ANN models with 2, 4, and 6 hidden layers with the Modified National Institute of Standards and Technology database (MNIST), it gave a maximum accuracy of 99.5% for handwritten digits classification than the results produced by the existing AFs. The proposed AF E-Tanh outperformed other existing and well-known AFs ReLU, Sigmoid, Tanh, Swish, and E-Swish. In deep ANNs with 5–50 hidden layers, the proposed AF's prediction accuracy of up to 25 hidden layers was higher than the existing AFs on the MNIST database. When used in shallow CNN for the CIFAR10 database, it outperformed other AFs. When used in wide residual network, it performed poorly in classifying images in CIFAR10 and CIFAR100 database but performed competitively with all other AFs in CNN models on the MNIST dataset. [ABSTRACT FROM AUTHOR]

Recently, many studies have been carried out on model compression to handle the high computational cost and high memory footprint brought by the implementation of deep neural networks. In this paper, model compression of convolutional neural networks is constructed as a multiobjective optimization problem with two conflicting objectives, reducing the model size and improving the performance. A novel structured pruning method called Conventional-based and Evolutionary Approaches Guided Multiobjective Pruning (CEA-MOP) is proposed to address this problem, where the power of conventional pruning methods is effectively exploited for the evolutionary process. A delicate balance in pruning rate and model accuracy has been automated achieved by a multiobjective optimization evolutionary model. First, an ensemble framework integrates pruning metrics to establish a codebook for further evolutionary operations. Then, an efficient coding method is developed to shorten the length of chromosome, thus ensuring its superior scalability. Finally, sensitivity analysis is automatically carried out to determine the upper bound of pruning rate for each layer. Notably, on CIFAR-10, CEA-MOP reduces more than 50% FLOPs on ResNet-110 and improves the relative accuracy. Moreover, on ImageNet, CEA-MOP reduces more than 50% FLOPs on ResNet-101 with negligible top-1 accuracy drop. [ABSTRACT FROM AUTHOR]

To elude from the camera, suspects often hide behind other things or persons, leading to a series of occlusion patterns. These suspects are notoriously hard to search due to the substantially various appearance in the intricate occlusion patterns. Existing methods solving occlusion problem depend on learning several frequent patterns separately. It brings not only high consumption but also less coverage of patterns in real application scenarios. Different from the current researches which only concern certain patterns that do not synthesize the occlusion patterns in practical applications, we consider a wide range of occlusion patterns which conform the real application scenarios in one coherent model with less interference of both the occlusion and background areas and without redundant computation. Consequently, we propose a channel-guided mechanism (CGM) for occluded suspect search in this paper. The core idea is that different body areas have been activated via different channels in convolutional neural networks. By suppressing the effects of the interference areas, such as occlusion and background areas, we can filter out the visible areas which are the essential elements for the occlusion patterns. Channel-aware attention is introduced to learn the relation between areas and channels. Furthermore, we can identify suspects using a rule which focuses more on the visible area and focuses less on the occluded area in the specific occlusion pattern. Extensive evaluations on two challenging datasets confirm the effectiveness of the proposed CGM. [ABSTRACT FROM AUTHOR]

Bridges play an important role in transportation, but because of overload and natural factors, bridges will inevitably be damaged, which will affect traffic and even lead to major accidents. Therefore, timely and accurate identification of bridge damage is extremely necessary. Because of the great danger of manual detection, in order to identify the damage of prestressed concrete girder bridge safely, conveniently and accurately, this paper proposes a method of damage identification of prestressed concrete girder bridge based on convolutional neural network, which realizes the intelligent identification of bridge damage. Firstly, the damage identification method based on the flexibility matrix is introduced, and the flexibility diagonal curvature index constructed by the diagonal element of flexibility matrix is introduced. Secondly, the basic principle of applying convolutional neural network to bridge damage identification is elaborated. Finally, combined with the flexibility curvature method and the convolutional neural network, the flexibility of the structure is selected as the input of the convolutional neural network to realize the bridge damage identification. Through simulation, it is found that the use of convolutional neural network for the bridge identification is feasible, and combined with the flexibility curvature method, it can well identify the damage location and damage degree of prestressed concrete beam bridge structure. [ABSTRACT FROM AUTHOR]

The affective/emotional status of patients is strongly connected to the healing process and their health. Therefore, being aware of the psychological peaks and troughs of a patient provides the advantage of timely intervention by specialists or closely related kinsfolk. In this context, this paper presents the design and implementation of an emotion analysis module integrated in an existing telemedicine platform. Two different methodologies are utilized and discussed. The first scheme exploits the fast and consistent properties of the speeded-up robust features algorithm in order to identify the existence of seven different sentiments in human faces. The second is based on convolutional neural networks. The whole functionality is provided as a Web service for the healthcare platform during regular video teleconference sessions between authorized medical personnel and patients. The paper discusses the technical details of the implementation and the incorporation of the proposed scheme and provides the initial results of its accuracy and operation in practice. [ABSTRACT FROM AUTHOR]

This paper introduces a new image-based handwritten historical digit dataset named Arkiv Digital Sweden (ARDIS). The images in ARDIS dataset are extracted from 15,000 Swedish church records which were written by different priests with various handwriting styles in the nineteenth and twentieth centuries. The constructed dataset consists of three single-digit datasets and one-digit string dataset. The digit string dataset includes 10,000 samples in red–green–blue color space, whereas the other datasets contain 7600 single-digit images in different color spaces. An extensive analysis of machine learning methods on several digit datasets is carried out. Additionally, correlation between ARDIS and existing digit datasets Modified National Institute of Standards and Technology (MNIST) and US Postal Service (USPS) is investigated. Experimental results show that machine learning algorithms, including deep learning methods, provide low recognition accuracy as they face difficulties when trained on existing datasets and tested on ARDIS dataset. Accordingly, convolutional neural network trained on MNIST and USPS and tested on ARDIS provide the highest accuracies 58.80 % and 35.44 % , respectively. Consequently, the results reveal that machine learning methods trained on existing datasets can have difficulties to recognize digits effectively on our dataset which proves that ARDIS dataset has unique characteristics. This dataset is publicly available for the research community to further advance handwritten digit recognition algorithms. [ABSTRACT FROM AUTHOR]

As a representative of shared mobility, bike sharing has become a green and convenient way to travel in cities in recent years. Bike usage prediction becomes more important for supporting efficient operation and management in bike share systems as the basis of inventory management and bike rebalancing. The essential of usage prediction in bike sharing systems is to model the spatial interactions of nearby stations, the temporal dependence of demands, and the impacts of environmental and societal factors. Deep learning has shown a great advantage of making a precise prediction for bike sharing usage. Recurrent neural networks capture the temporal dependence with the memory cell and gate mechanisms. Convolutional neural networks and graph neural networks learn spatial interactions of nearby stations with local convolutional operations defined for the grid-format and graph-format inputs respectively. In this survey, the latest studies about bike sharing usage prediction with deep learning are reviewed, with a classification for the prediction problems and models. Different applications based on bike usage prediction are discussed, both within and beyond bike share systems. Some research directions are pointed out to encourage future research. To the best of our knowledge, this paper is the first comprehensive survey that focuses on bike sharing usage prediction with deep learning techniques. [ABSTRACT FROM AUTHOR]

Pressure injury follow-up and treatment is a very costly and significant health care problem for many countries. Early and accurate diagnosis and treatment planning are critical for effective treatment of pressure injuries. Interventional information retrieval methods are both painful for patients and increase the risk of infection. However, thanks to non-invasive techniques such as imaging systems, it is possible to monitor pressure wounds more easily without causing any harm to patients. The purpose of this research is to develop a deep learning-based system for the analysis and monitoring of pressure injuries that provides an automatic classification of pressure injury stages. This paper introduces the pressure injury images dataset (PIID): a novel dataset for the classification of pressure injuries stages. We hope that PIID will encourage further research on the automatic visual classification of pressure injury stages. We also perform extensive analyses on PIID using state-the-of-art convolutional neural networks architectures with the power of transfer learning and image augmentation techniques. [ABSTRACT FROM AUTHOR]

Image fusion can provide more comprehensive information by integrating images from different sensors. So far, intensive research has been devoted to utilizing deep learning to fuse images from multiple sensors more effectively. However, most existing image fusion methods are single-scene-specific with substantial model complexity, and none of the current deep learning-based methods has the noise suppression capability. In this paper, we propose a simple and effective unified image fusion framework with noise suppression capability. The core of the proposed method is a novel self-supervised image reconstruction task. Specifically, we destroy several selected regions of an image in the frequency domain by utilizing discrete wavelet transform and reconstruct it by an encoder–decoder model. Then, in the fusion phase, the two images to be fused are first encoded by the encoder and the features are averaged and then decoded to obtain the final fused image. To keep the fusion network small, we construct the encoder–decoder model with three-layer convolutional neural network. In this way, the fusion phase is much faster than the latest works (only 0.0082s on average). We conducted experiments on four image fusion scenarios and synthesized a new experiment of fusing noisy images. Both subjective and objective evaluations demonstrate that the proposed method outperforms existing state-of-the-art approaches in these experiments. Codes are available at https://github.com/slliuEric/WaveSSL. [ABSTRACT FROM AUTHOR]

Monitoring the perceptual quality of digital images is fundamentally important since digital image transmissions through the Internet continue to increase exponentially. Many automatic image quality evaluation (IQE) metrics have been developed based on image features correlated to image distortions; however, those metrics are only effective on particular image distortion types. In recent years, convolutional neural network (CNNs) have been developed for IQEs. These CNNs first capture image features from distorted images; image qualities are predicted based on the captured image features. Since the CNN weights are randomly initialized and are updated with respect to a loss function, image features which are strongly correlated to image quality are not guaranteed to be captured. In this paper, a hybrid deep neural network (DNN) is proposed by integrating image quality metrics to capture image features which are correlated to image quality; the approach guarantees that significant image features are included to predict image quality. Also, a tree-based classifier namely geometric semantic genetic programming is proposed to perform the overall predictions by incorporating CNN predictions and image features; the approach is simpler than the fully connected network but is able to model the nonlinear image qualities. The performance of the proposed hybrid DNN is evaluated by an image quality database with 3000 distorted images. The mean correlation achieved by the proposed hybrid DNN is 0.57 which is higher than the other tested methods. Experimental results with the t- test, F-test and Tueky's range tests show that the proposed hybrid DNN achieves more accurate image predictions with a 99.9% confidence level, compared to the state-of-the-art IQE metics and the most recently developed CNN for IQEs. [ABSTRACT FROM AUTHOR]

In recent years, the Internet of Things (IoT) has received a lot of attention. It has been used in many applications such as the control industry, industrial plants, and medicine. In this regard, a fundamental necessity is to implement security in IoT. To this end, Network intrusion detection systems (NIDSs) have been recently in the detection of network attacks and threats. Currently, these systems use a variety of deep learning (DL) models such as the convolutional neural networks to improve the detection of attacks. However, almost all current DL-based NIDSs are made up of many layers, and therefore, they need a lot of processing resources because of their high number of parameters. On the other hand, due to the lack of processing resources, such inefficient DL models are unusable in IoT devices. This paper presents a very accurate NIDS that is named DFE, and it uses a very lightweight and efficient neural network based on the idea of deep feature extraction. In this model, the input vector of the network is permuted in a 3D space, and its individual values are brought close together. This allows the model to extract highly discriminative features using a small number of layers without the need to use large 2D or 3D convolution filters. As a result, the network can achieve an accurate classification using a significantly small number of required calculations. This makes the DFE ideal for real-time intrusion detection by IoT devices with limited processing capabilities. The efficacy of the DFE has been evaluated using three popular public datasets named UNSW-NB15, CICIDS2017, and KDDCup99, and the results show the superiority of the proposed model over the state-of-the-art algorithms [ABSTRACT FROM AUTHOR]

It is still hard to deal with artifacts in magnetic resonance images (MRIs), particularly when the latter are to be segmented. This paper introduces a novel deep-based scheme for tumor segmentation in brain MRIs. According to the proposed scheme, a large set of partial sub-images are paved form sliced from an MRI volume then inputted to an ensemble of convolutional neural networks (CNNs) in order to label the voxels in the centers of the sub-images, according to the classes to which they should belong. Partial sub-images, that capture local patterns around central voxels, have allowed to speed-up both the training and the prediction steps, allowing efficient use of such a scheme for real MRI-based tumor diagnosis. Experiments were performed using the BraTS (brain tumor segmentation) database, where the obtained results show that the proposed scheme allows both fast and accurate brain tumor detection and segmentation in pathological MRIs. [ABSTRACT FROM AUTHOR]

In the financial market, the stock price prediction is a challenging task which is influenced by many factors. These factors include economic change, politics and global events that are usually recorded in text format, such as the daily news. Therefore, we assume that real-world text information can be used to forecast stock market activity. However, only a few works considered both text and numerical information to predict or analyse stock trends. These works used preprocessed text features as the model inputs; therefore, latent information in text may be lost because the relationships between the text and stock price are not considered. In this paper, we propose a fusion network, i.e. a spatial-temporal attention-based convolutional network (STACN) that can leverage the advantages of an attention mechanism, a convolutional neural network and long short-term memory to extract text and numerical information for stock price prediction. Benefiting from the utilisation of an attention mechanism, reliable text features that are highly relevant to stock value can be extracted, which improves the overall model performance. The experimental results on real-world stock data demonstrate that our STACN model and training scheme can handle both text and numerical data and achieve high accuracy on stock regression tasks. The STACN is compared with CNNs and LSTMs with different settings, e.g. a CNN with only stock data, a CNN with only news titles and LSTMs with only stock data. CNNs considering only stock data and news titles have mean squared errors of 28.3935 and 0.1814, respectively. The accuracy of LSTMs is 0.0763. The STACN can achieve an accuracy of 0.0304, outperforming CNNs and LSTMs in stock regression tasks. [ABSTRACT FROM AUTHOR]

Recently, neural architecture search (NAS) has gained a lot of attention as a tool for constructing deep neural networks automatically. NAS methods have successfully found convolutional neural networks (CNNs) that exceed human expert-designed networks on image classification in computer vision. However, there are growing demands for semantic segmentation in several areas including remote sensing image analysis. In this paper, we introduce an evolutionary NAS method for semantic segmentation of high-resolution aerial images. The proposed method leverages the complementary strengths of gene expression programming and cellular encoding to develop an encoding scheme, called symbolic linear generative encoding (SLGE), for evolving cells (directed acyclic graphs) as building-blocks to construct modularized encoder-decoder CNNs via an evolutionary process. SLGE can evolve cells with multi-branch and shortcut connections similar to the Inception-ResNet-like modules which can improve training and inference performance in deep neural networks. In experiments, we demonstrate the effectiveness of the proposed method on the challenging ISPRS Vaihingen, Potsdam and UAVid semantic segmentation benchmarks. Compared with recent state-of-the-art systems, our network, dubbed SLGENet, improves the overall accuracy performance on Vaihingen and Potsdam; and achieves a competitive overall accuracy on UAVid using fewer parameters. Our method achieves promising results in a little time of 2.5 GPU days. [ABSTRACT FROM AUTHOR]

Multimedia IoT (M-IoT) is an emerging type of Internet of things (IoT) relaying multimedia data (images, videos, audio and speech, etc.). The rapid growth of M-IoT devices enables the creation of a massive volume of multimedia data with different characteristics and requirements. With the development of artificial intelligence (AI), AI-based multimedia IoT systems have been recently designed and deployed for various video-based services for contemporary daily life, like video surveillance with high definition (HD) and ultra-high definition (UHD) and mobile multimedia streaming. These new services need higher video quality in order to meet the quality of experience (QoE) required by the users. Versatile video coding (VVC) is the new video coding standard that achieves significant coding efficiency over its predecessor high-efficiency video coding (HEVC). Moreover, VVC can achieve up to 30% BD rate savings compared to HEVC. Inspired by the rapid advancements in deep learning, we propose in this paper a wide-activated squeeze-and-excitation deep convolutional neural network (WSE-DCNN) technique-based video quality enhancement for VVC. Therefore, we replace the conventional in-loop filtering in VVC by the proposed WSE-DCNN model that eliminates the compression artifacts in order to improve visual quality and hence increase the end user QoE. The obtained results prove that the proposed in-loop filtering technique achieves - 2.85 %, - 8.89 %, and - 10.05 % BD rate reduction for luma and both chroma components under random access configuration. Compared to the traditional CNN-based filtering approaches, the proposed WSE-DCNN-based in-loop filtering framework achieves efficient performance in terms of RD cost. [ABSTRACT FROM AUTHOR]

COVID-19 has taken a toll on the entire world, rendering serious illness and high mortality rate. In the present day, when the globe is hit by a pandemic, those suspected to be infected by the virus need to confirm its presence to seek immediate medical attention to avoid adverse outcomes and also to prevent further transmission of the virus in their close contacts by ensuring timely isolation. The most reliable laboratory testing currently available is the reverse transcription–polymerase chain reaction (RT-PCR) test. Although the test is considered gold standard, 20–25% of results can still be false negatives, which has lately led physicians to recommend medical imaging in specific cases. Our research examines the aspect of chest imaging as a method to diagnose COVID-19. This work is not directed to establish an alternative to RT-PCR, but to aid physicians in determining the presence of virus in medical images. As the disease presents lung involvement, it provides a basis to explore computer vision for classification in radiographic images. In this paper, authors compare the performance of various models, namely ResNet-50, EfficientNetB0, VGG-16 and a custom convolutional neural network (CNN) for detecting the presence of virus in chest computed tomography (CT) scan and chest X-ray images. The most promising results have been derived by using ResNet-50 on CT scans with an accuracy of 98.9% and ResNet-50 on X-rays with an accuracy of 98.7%, which offer an opportunity to further explore these methods for prospective use. [ABSTRACT FROM AUTHOR]

High-impedance faults (HIFs) exhibit low current amplitude and highly diverse characteristics, which make them difficult to be detected by conventional overcurrent relays. Various machine learning (ML) techniques have been proposed to detect and classify HIFs; however, these approaches are not reliable in presence of diverse HIF and non-HIF conditions and, moreover, rely on resource-intensive signal processing techniques. Consequently, this paper proposes a novel HIF detection and classification approach based on a state-of-the-art deep learning model, the transformer network, stacked with the Convolutional neural network (CNN). While the transformer network learns the complex HIF pattern in the data, the CNN enhances the generalization to provide robustness against noise. A kurtosis analysis is employed to prevent false detection of non-fault disturbances (e.g., capacitor and load switching) and nonlinear loads as HIFs. The performance of the proposed HIF detection and classification approach is evaluated using the IEEE 13-node test feeder. The results demonstrate that the proposed protection method reliably detects and classifies HIFs, is robust against noise, and outperforms the state-of-the-art techniques. [ABSTRACT FROM AUTHOR]

Financial time series are chaotic that, in turn, leads their predictability to be complex and challenging. This paper presents a novel financial time series prediction hybrid that involves Chaos Theory, Convolutional neural network (CNN), and Polynomial Regression (PR). The financial time series is first checked in this hybrid for the presence of chaos. The chaos in the series of times is later modeled using Chaos Theory. The modeled time series is input to CNN to obtain initial predictions. The error series obtained from CNN predictions is fit by PR to get error predictions. The error predictions and initial predictions from CNN are added to obtain the final predictions of the hybrid model. The effectiveness of the proposed hybrid (Chaos+CNN+PR) is tested by using three types of Foreign exchange rates of financial time series (INR/USD, JPY/USD, SGD/USD), commodity prices (Gold, Crude Oil, Soya beans), and stock market indices (S&P 500, Nifty 50, Shanghai Composite). The proposed hybrid is superior to Auto-regressive integrated moving averages (ARIMA), Prophet, Classification and Regression Tree (CART), Random Forest (RF), CNN, Chaos+CART, Chaos+RF and Chaos+CNN in terms of MSE, MAPE, Dstat, and Theil's U. [ABSTRACT FROM AUTHOR]

Drogue detection is a crucial component for autonomous aerial refueling. Drogue detection requires robust detection of the drogue in real time and different environments. A real-time embedded drogue detection method based on quantized convolutional neural networks is proposed in this paper to cope with this issue. A backbone network that is based on a multi-receptive field stem and multi-receptive field dense blocks are designed to quickly extract the features of images. An optimized detection network that is based on a depth-wise separable residual prediction block is used to achieve drogue detection in real time. A real aerial refueling dataset was built for training network model. A ground test and video sequences experiment were conducted to demonstrate the effectiveness, robustness, and real time of the proposed approach. [ABSTRACT FROM AUTHOR]

In recent studies, machine learning and deep learning strategies have been explored in many EEG-based application for best performance. More specifically, convolutional neural networks (CNNs) have demonstrated incredible capacity in electroencephalograph (EEG)-evoked emotion classification tasks. In preexisting case, CNN-based emotion classification techniques using EEG signals mostly involve a moderately intricate phase of feature extrication before any network model implementation. The CNNs are not able to well describe the natural interrelation among the various EEG channels, which basically provides essential data for the classification of different emotion states. In this paper, an efficacious and advanced version of CNN called Emotion-based Capsule Network (EmotionCapsNet) for multi-channel EEG-based emotion classification to achieve better classification accuracy is presented. EmotionCapsNet has been applied to the raw EEG signals as well as 2D image representation generated from EEG signals which can extricate descriptive and complex features from the EEG signals and decide the different emotional states. The proposed system is then compared with the other conventional machine learning and deep learning-based CNN model. Our strategy accomplishes an average accuracy of 77.50%, 78.44% and 79.38% for valence, arousal and dominance on the DEAP, 79.06%, 78.90% and 79.69% on AMIGOS and attains an average accuracy of 80.34%, 83.04% and 82.50% for valence, arousal and dominance on the DREAMER, respectively. These outcomes demonstrate that adapted strategy yields comparable precision on raw EEG signal and it also provides better classification results on spatiotemporal feature of EEG signal for emotion classification task. [ABSTRACT FROM AUTHOR]

The redundancy in convolutional neural networks (CNNs) causes a significant number of extra parameters resulting in increased computation and less diverse filters. In this paper, we introduce filter pruning via expectation-maximization (FPEM) to trim redundant structures and improve the diversity of remaining structures. Our method is designed based on the discovery that the filter diversity of pruned networks is positively correlated with its performance. The expectation step divides filters into groups by maximum likelihood layer-wisely, and averages the output feature maps for each cluster. The maximization step calculates the likelihood estimation of clusters and formulates a loss function to make the distributions in the same cluster consistent. After training, the intra-cluster redundant filters can be trimmed and only intra-cluster diverse filters are retained. Experiments conducted on CIFAR-10 have outperformed the corresponding full models. On ImageNet ILSVRC12, FPEM reduces 46.5 % FLOPs on ResNet-50 with only 0.36 % Top-1 accuracy decrease, which advances the state-of-arts. In particular, the FPEM offers strong generalization performance on the object detection task. [ABSTRACT FROM AUTHOR]

In Genome-Wide Association Studies (GWAS), detection of T2D-related variants in genome sequences and accurate modeling of the complex structure of the relevant gene are of great importance for the diagnosis of diabetes. For this purpose, this paper presents a novel strong algorithm to accurately and effectively identify Type 2 Diabetes (T2D) risk variants at high-performance rates. The proposed algorithm consists of five important phases. The first stage is to collect T2D-associated DNA sequences and to digitize them by the Entropy-based technique. The second stage is to transform these digitized DNA sequences into 224 × 224 pixels size spectrum images. The third is to extract a distinctive feature set from these spectrum images using the ResNet and VGG19 architectures. The fourth is to classify the effective feature set using SVM and k-NN methods. The last stage is to evaluate the system with k-fold cross-validation. As a result of the developed algorithm, the performances of the used Convolutional Neural Network (CNN) methods, the Entropy-based technique, and the classifiers were compared in relation. As a result of the study a combination model of the proposed Entropy-based technique, ResNet and Support Vector Machine (SVM) achieved the highest accuracy rate with 99.09%. With this study, the performance of the system in the extraction of epigenetic features and prediction of T2D from spectrogram images was investigated. The results show that the system will contribute to the identification of all genes in diabetes-related tissue and studies on new drug targets. [ABSTRACT FROM AUTHOR]

Despite some recent achievements in the intelligent data-driven fault diagnosis of mechanical rotating components, in most cases very large amounts of pure data are demanded for model training. However, the collected vibration signals of mechanical rotating components are inevitably contaminated with noise in real industries. To resolve the low fault diagnosis accuracy due to strong noise interference, a convolutional neural network with interference suppression (ISCNN) is proposed in this paper. First, a parallel convolutional structure with dilated wide convolution kernels is designed to extract long-time correlated fault features of multiple scales from a noise-contaminated signal by sparse sampling, and noise interference is suppressed by filtering operations. Then, a multiscale feature enhancement module is constructed to achieve adaptive tuning of time scale by employing a parallel selective kernel network to exploit weak fault features buried in noisy signals. Finally, a convolutional feature fusion method is adopted to integrate multidimensional fault features and feed them into the classifier, thus achieving accurate fault diagnosis of mechanical rotating components. Experimental results on two benchmark datasets indicate that the ISCNN model outperforms its counterparts in the fault diagnosis field and improves the fault discriminability of mechanical rotating components in strong noisy scenarios. [ABSTRACT FROM AUTHOR]

Spectrograms provide rich feature information of music data. Significant progress has been made in music classification using spectrograms and Convolutional Neural Networks (CNNs). However, the softmax loss commonly used in existing CNNs lacks sufficient power to discriminate deep features of music. To overcome this limitation, we propose a Combined Angular Margin and Cosine Margin Softmax Loss (AMCM-Softmax) approach in this paper to enhance intra-class compactness and inter-class discrepancy simultaneously. Specifically, normalization on the weight vectors and feature vectors is adopted to eliminate radial variations. Then, an angular margin parameter and a cosine margin parameter are introduced to maximize the decision margin by enforcing angular and cosine margin constraints. Consequently, the discrimination of features is enhanced by normalization and margin maximization. The decision boundary and the target logit curve of AMCM-Softmax can provide a clear geometric interpretation. Extensive experiments on music datasets show that AMCM-Softmax consistently outperforms the current state-of-the-art approaches in classifying genre and emotion. Our work also shows that a margin loss function can lead to better performance and be used in an advanced CNN model for music classification. [ABSTRACT FROM AUTHOR]

Convolutional neural networks have pushed forward image analysis research and computer vision over the last decade, constituting a state-of-the-art approach in object detection today. The design of increasingly deeper and wider architectures has made it possible to achieve unprecedented levels of detection accuracy, albeit at the cost of both a dramatic computational burden and a large memory footprint. In such a context, cloud systems have become a mainstream technological solution due to their tremendous scalability, providing researchers and practitioners with virtually unlimited resources. However, these resources are typically made available as remote services, requiring communication over the network to be accessed, thus compromising the speed of response, availability, and security of the implemented solution. In view of these limitations, the on-device paradigm has emerged as a recent yet widely explored alternative, pursuing more compact and efficient networks to ultimately enable the execution of the derived models directly on resource-constrained client devices. This study provides an up-to-date review of the more relevant scientific research carried out in this vein, circumscribed to the object detection problem. In particular, the paper contributes to the field with a comprehensive architectural overview of both the existing lightweight object detection frameworks targeted to mobile and embedded devices, and the underlying convolutional neural networks that make up their internal structure. More specifically, it addresses the main structural-level strategies used for conceiving the various components of a detection pipeline (i.e., backbone, neck, and head), as well as the most salient techniques proposed for adapting such structures and the resulting architectures to more austere deployment environments. Finally, the study concludes with a discussion of the specific challenges and next steps to be taken to move toward a more convenient accuracy–speed trade-off. [ABSTRACT FROM AUTHOR]

Cloud computing enables clients to acquire cloud resources dynamically and on demand for their cloud applications and services. For cloud providers, especially, Software as a Service (SaaS) providers, the prediction of future cloud resource requirements, such as CPU usage for their cloud applications, to implement client requests is a complex task because it depends on incoming workloads. Due to workload fluctuations, it is difficult for SaaS cloud providers to predict or forecast future demand for resource usage in the next time interval and, accordingly, to allocate the required resources. Furthermore, cloud computing systems consist of many virtual machines (VMs), which increases the complexity of the prediction problem due to the correlations that exist between the large workload data in these VMs. Therefore, accurate resource usage forecasting remains a challenge, and relatively few studies have explored the prediction of CPU usage for VMs in cloud data centers. This paper proposes an autonomic and intelligent workload forecasting method for cloud resource provisioning based on the concept of autonomic computing and a deep learning approach. In particular, to predict future demand for CPU usage and determine how to respond to workload fluctuations in the next interval, we propose an efficient deep learning model based on a diffusion convolutional recurrent neural network (DCRNN). Existing deep learning models that are widely applied cannot handle accurate real-time forecasting due to the presence of inconsistent and nonlinear workloads in cloud computing systems. The goal of the proposed deep learning model is to improve forecasting accuracy and minimize the error between the predicted and the actual workloads. The effectiveness of the proposed DCRNN-based deep learning model was evaluated using experiments on a real-world dataset of PlanetLab's CPU usage traces. The results indicate that the proposed approach outperformed other existing deep learning models, achieving a mean absolute percentage error of 0.18 and root-mean-square error of 2.40. [ABSTRACT FROM AUTHOR]

This paper presents a novel method for improving the invariance of convolutional neural networks (CNNs) to selected geometric transformations in order to obtain more efficient image classifiers. A common strategy employed to achieve this aim is to train the network using data augmentation. Such a method alone, however, increases the complexity of the neural network model, as any change in the rotation or size of the input image results in the activation of different CNN feature maps. This problem can be resolved by the proposed novel convolutional neural network models with geometric transformations embedded into the network architecture. The evaluation of the proposed CNN model is performed on the image classification task with the use of diverse representative data sets. The CNN models with embedded geometric transformations are compared to those without the transformations, using different data augmentation setups. As the compared approaches use the same amount of memory to store the parameters, the improved classification score means that the proposed architecture is more optimal. [ABSTRACT FROM AUTHOR]

Human action recognition is a process of labeling video frames with action labels. It is a challenging research topic since the background of videos is usually chaotic, which will reduce the performance of traditional human action recognition methods. In this paper, we propose a novel spatiotemporal saliency-based multi-stream ResNets (STS), which combines three streams (i.e., a spatial stream, a temporal stream and a spatiotemporal saliency stream) for human action recognition. Further, we propose a novel spatiotemporal saliency-based multi-stream ResNets with attention-aware long short-term memory (STS-ALSTM) network. The proposed STS-ALSTM model combines deep convolutional neural network (CNN) feature extractors with three attention-aware LSTMs to capture the temporal long-term dependency relationships between consecutive video frames, optical flow frames or spatiotemporal saliency frames. Experimental results on UCF-101 and HMDB-51 datasets demonstrate that our proposed STS method and STS-ALSTM model obtain competitive performance compared with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]