Search

Actual wind farms usually perform condition monitoring of wind turbines based on simple trigger logic, which often have high false alarm rates and overlapping warnings. Supervisory control and data acquisition (SCADA) data contain operational information of wind turbines, based on which automatic fault diagnosis is extremely valuable. However, SCADA data have complex characteristics such as high-dimensional variable correlation, dynamic variability, and temporal correlation which bring challenges to fault diagnosis. Therefore, this paper proposes a novel fault diagnosis method to automatically identify the operating state of wind turbines: adaptive multivariate time-series convolutional network (AdaMTCN). AdaMTCN resamples SCADA data with multiple time steps at first, and the obtained resampling matrices contain multivariate time series information of different states. Then, multivariate time series convolutional networks (MTCN) are innovatively proposed, which extract enriched features from the multiple time step resampling matrices. Finally, according to the proposed adaptive decision fusion method, multiple MTCN models are weighted and fused into the overall AdaMTCN. Experimental results based on real case show that the AdaMTCN has excellent diagnostic performance in the face of SCADA data with complex characteristics. [ABSTRACT FROM AUTHOR]

Non-technical losses (NTL) in power grid not only cause significant financial loss but also reduce the quality of supply for the distribution network operators (DNOs). In recent years, the frequency of power theft activities related to NTL have been increased rapidly. The data-based detection approach for NTLs has been studied by researchers from all over the world thoroughly, but the region domain shift issues and the lack of labelled data limited its applicability. In this paper, an adversarial learning-based approach is proposed to alleviate the domain shift problem, named Cross-Regional Adaptive Network (CAN). First of all, a multilayer Convolutional Neural Networks (CNN) feature extractor maps the training and application scenario data into the same feature space. After that the feature extractor and the domain classifier are trained against each other by inversing extractor's gradients. Finally, the training and application scenario data are indistinguishable for the domain classifier, and the feature extractor can extract domain invariant features. Experiments are conducted on two sets of electricity consumption data captured by utility company to verify the effectiveness of the proposed method. • It is found that the DA-EFD algorithm with domain adaptation improved 23.1%, 13.9% and 14.0% in accuracy, AUC and PR-AUC over EFD without domain adaptation. • DA-EFD algorithm can transfer model trained on labelled datasets to unlabelled data without any labelled information in the target domain, which may help to address the scarcity of labelled data. [ABSTRACT FROM AUTHOR]

With the growth of global energy demand, the proportion of the total installed wind capacity continues to increase. Wind speed forecasting is essential to enhance the utilization of wind energy. However, it is not easy to make accurate wind speed forecasting since wind speed time series data has nonlinearity, fluctuation, and intermittence. These features of wind speed time series data make classical models unable to obtain expected forecasting results. Many researchers proposed wind speed forecasting models to overcome the shortage. Nevertheless, some wind speed prediction methods are limited because they cannot mine all implicit features of wind speed data. This paper proposes a novel wind speed forecasting model that can capture all implicit information of wind speed data and obtain accuracy prediction results. To be specific, ensemble empirical mode decomposition (EEMD) is firstly applied to remove the noise in the original wind speed data. Secondly, we build a novel deep learning model based on the attention mechanism and convolutional neural network (CNN). The proposed forecasting model can focus the important part of wind speed data and reduce the high computational complexity of CNN. Finally, a full connect neural network layer is employed to obtain wind speed forecasting results. In order to validate the performance of the proposed model, we take the wind speed data of the M2 tower at 20-meter height of the National Wind Power Technology Center of the United States as an example. The experimental results demonstrate that the forecasting errors of the proposed model are smaller than other comparative models. And the Diebold-Mariano test confirms that the proposed model exhibits a significant difference in performance compared with the comparison models. [ABSTRACT FROM AUTHOR]

Automatic voice pathology detection using non-invasive techniques that utilize patients' speech and electroglottograph (EGG) signals play a vital role in diagnosis and early medical intervention. In this paper, a novel deep Multi-Modal and Multi-Layer Hybrid Fusion Network (MMHFNet) is proposed to improve the performance of non-invasive voice pathology detection systems. MMHFNet simultaneously incorporates complementary information of different modalities (speech and EGG signals). It also vertically combines the low-level features, extracted from shallow layers, and high-level features, extracted from deep layers, to take the full advantage of spatio-spectral information of different layers for multi-layer fusion. The features extracted by MMHFNet are then fed into an LSTM classification network to diagnose the voice pathology. Comprehensive experiments are conducted on the publicly available Saarbruecken Voice Database (SVD) to evaluate the performance of the proposed MMHFNet. This dataset is used in two manners; one using its all samples and the other with selected samples to form the largest balanced SVD dataset. Experimental results demonstrated that the proposed MMHFNet achieves accuracy rates of 91% and 96.05% for datasets with all and balanced samples, respectively. • Novel multi-modal/layer fusion framework is proposed for voice pathology detection. • Multi-modal data fusion of speech and EGG signals is employed. • Multi-layer fusion is utilized using deep and shallow layers' features. • Performance of the proposed network is evaluated on the challenging SVD Dataset. • Experimental results demonstrated a new state-of-the-art performance. [ABSTRACT FROM AUTHOR]

Orthogonal frequency division multiplexing (OFDM) the signal processing is a key issue in wireless communication research. The multipath effect and Doppler shift of wireless communication channels can lead to distortion of the transmitted signal, which poses a considerable challenge to the information recovery of communication receivers. This paper presents the signal processing method of OFDM communication based on convolutional neural network (CNN). The method replaces all signal processing modules of the OFDM communication receiver with CNN, and the information is recovered by the CNN. In order to adapt to the processing of communication signals, we designed a one-dimensional convolutional neural network (1D-CONV-CNN) model as the neural network structures by this method. Simulation results indicate that the signal processing method effectively reduces the bit error rate (BER) and improves its performance compared with the conventional reception method under different channel conditions. [ABSTRACT FROM AUTHOR]

Numerous studies have investigated image classification. In particular, recent methods based on deep learning have exhibited high accuracies. However, various existing state-of-the-art methods based on deep learning show different accuracies depending on the database and environment. Accordingly, different deep learning models need to be used in image classification studies according to the database, environment, and research field. This study investigated a technique to increase the accuracy of the existing deep learning-based models. The proposed method was applied to various existing state-of-the-art methods. In the proposed method, a convolution neural network (CNN) is trained using the classification activation map (CAM) to focus on specific areas in the input image. The CAM image is used as the ground-truth image. Furthermore, the concept of the CAM-based categorical adversarial network (CAM-CAN), in which the CNN is trained based on a generative adversarial network, is proposed in this paper. An action recognition experiment was performed using the self-collected Dongguk thermal image database (DTh-DB) and open database, and the results revealed that the accuracies of the existing state-of-the-art methods significantly increased after applying the proposed method. For instance, the accuracies obtained using the DTh-DB, TPR, PPV, ACC, and F1 with the conventional DenseNet201 model were 80.14%, 75.28%, 96.0%, and 75.91%, respectively. After applying the proposed method, the accuracies increased to 86.53%, 89.90%, 97.64%, and 85.84%, respectively. [ABSTRACT FROM AUTHOR]

• LSTM, WaveNet and CNN are developed to predict nonlinear structural responses. • SDOF, steel building frame and bridge structures are used to test these models. • Sensitivity to different training samples and model architectures are analyzed. Predicting the nonlinear time-history responses of civil engineering structures under seismic loading remains an essential task in earthquake engineering. This paper explores the promise of developing three deep learning (DL) models, i.e., long short-term memory (LSTM), WaveNet, and 2D convolution neural network (CNN), to predict seismic response time histories of three benchmark structures, including a single degree-of-freedom (SDOF) system, a steel building frame, and a multi-component bridge structure. LSTM has been previously developed and is utilized to serve as a reference model, while WaveNet and 2D CNN (i.e., it deals with the data in coupled time–frequency dimensions) are newly developed in the current study. One other novel contribution is to replace the final layer of the WaveNet with an LSTM layer, which significantly improves the model performance. Methodological backgrounds of these DL models are introduced, followed by discussions on model architectures and hyperparameters, the list of evaluation metrics, and the ground motion (GM) suite selected for data generation. High-fidelity numerical models are developed for conducting nonlinear time history analyses that generate numerous motion-response pairs for training, validating, and testing the DL models. The LSTM and WaveNet directly use time series GM inputs and seismic response outputs to train the models, whereas the CNN makes inferences on time–frequency spectrogram images converted through the short-time Fourier transform (STFT). These three DL models are investigated by comparing deterministic predictions under one testing GM and probabilistic distributions of six evaluation metrics. The models' accuracy, efficiency, and robustness are further examined using a sensitivity study under different training samples and model architectures. Research findings from this study provide a sound reference for the community to leverage these three DL models to achieve reliable and efficient time history response predictions, which are crucial for fulfilling cutting-edge research/practical tasks such as regional seismic risk assessment and performance-based seismic design. [ABSTRACT FROM AUTHOR]

Automatic Visual Captioning (AVC) generates syntactically and semantically correct sentences by describing important objects, attributes, and their relationships with each other. It is classified into two categories: image captioning and video captioning. It is widely used in various applications such as assistance for the visually impaired, human-robot interaction, video surveillance systems, scene understanding, etc. With the unprecedented success of deep-learning in Computer Vision and Natural Language Processing, the past few years have seen a surge of research in this domain. In this survey, the state-of-the-art is classified based on how they conceptualize the captioning problem, viz., traditional approaches that cast visual description either as retrieval or template-based description and deep learning approaches. A detailed review of existing methods, highlighting their pros and cons, societal impact as the number of citations, architectures used, datasets experimented on and GitHub link is presented. Moreover, the survey also provides an overview of the benchmark image and video datasets and the evaluation measures that have been developed to assess the quality of machine-generated captions. It is observed that dense or paragraph generation and Change Image Captioning (CIC) are stimulating the research community more due to the near-to-human abstraction ability. Finally, the paper explores future directions in the area of automatic visual caption generation. [ABSTRACT FROM AUTHOR]

As a popular method for modeling violent free surface flow, the incompressible Smoothed Particle Hydrodynamics (ISPH) based on the Lagrangian formulation has attracted a great attention worldwide. The Lagrangian ISPH solves the unsteady Navier-Stokes and continuity equations using the projection method, in which the pressure is obtained by solving the pressure Poisson's equation (PPE) that is the most time-consuming part in the ISPH procedure. In this paper, the Convolutional Neural Network (CNN) is combined with ISPH and used to predict the fluid pressure instead of solving the PPE directly. Although limited attempts of using CNN for solving the PPE in Eulerian formulation (referred to as the Eulerian CNN framework) in mesh-based methods are found in the public domain, the present model is the first ISPH model supported by CNN in a Lagrangian formulation. The proposed model overcome several challenges associated with combining CNN with ISPH, including selecting the input parameters, formulating the objective functions, producing the training dataset and dealing with boundary conditions. Two classic free surface problems, i.e. the dam breaking and the wave propagation, are simulated to evaluate the performance of the present model. Quantitative assessments of the numerical error in terms of both the free surface profile and the pressure field are carried out. The assessments show that the new model does not only give results with satisfactory accuracy, but also requires much less computation time for estimating pressure if the number of particles is large, e.g., 100 thousands particles that is usually required in the practical ISPH simulation for free surface flow. [ABSTRACT FROM AUTHOR]

Aiming at the limitation of Hamiltonian neural network in high-dimensional data processing with non-observable physical quantity system, an embedded Hamiltonian dynamic evolutionary neural network model (EHDN) is proposed in this paper. First, the dynamic physical quantities of high dimensional nonlinear systems are represented by constructing Hamiltonian functions. Then a symplectic integrator is designed to realize the time-varying evolution of network features based on Hamiltonian regular equation and variational numerical integration. Finally, the Hamiltonian dynamic evolutionary neural network is embedded in the convolutional neural network (CNN) to realize the Hamiltonian simulation evolution of convolutional features. Experimental results on two different datasets, CIFAR and Fashion-Mnist, show that compared with existing state-of-the-art (SOTA) methods, EHDN model can combine the advantages of Hamiltonian dynamics and convolutional network to improve the recognition accuracy and training stability in high-dimensional image classification tasks. • An embeddable Hamiltonian neural network model is proposed to solve the degradation problem of very deep networks. • The high-dimensional image features are self-evolved by the latent Hamiltonian to reduce the hyperparametric constraints. • EHDN network architecture is flexible, and two evolution modes can be designed on demand. [ABSTRACT FROM AUTHOR]

Carbon emission monitoring is the key to achieving global emission reduction. This paper proposes a carbon emission identification method based on remote sensing images. Based on the regional satellite remote sensing image data set, the convolutional neural network model is used to extract the carbon emission characteristics of the region. The multi-model fully connected layer fusion algorithm and gamma correction data augmentation strategy are designed to analyze the accuracy of carbon emission monitoring. The results show that the results will be different when the image can not be preprocessed. In the case of Gamma correction, the model's accuracy is affected, but the robustness of the model is improved. The single-model local image cascade fusion algorithm has higher environmental adaptability than the single-model non-cascade algorithm. The recognition accuracy of the multi-model fusion algorithm is more than 7% higher than that of the single-model local image cascade fusion algorithm. The fully connected layer fusion accuracy is more than 1% higher than that of the feature map fusion, and its recognition accuracy for the data set reaches 87.21% and 88.35%, respectively. [ABSTRACT FROM AUTHOR]

• CNN based joint transmit and receive antenna selection. • Norm based antenna selection for MIMO VANETs. • Capacity enhancement of MIMO VANET by low complexity antenna selection algorithm. In this paper, we propose deep learning based transmit and receive antenna selection scheme for multiple-input-multiple-output vehicular adhoc networks (MIMO-VANETs). The dynamic nature of VANETs instills the need for regular updates in antenna selections, moreover massive computations involved for antenna selection due to changes in channel statistics results in large delay in message communication. The delay in antenna selection in case of conventional approaches can be reduced substantially by automating the redundant computations using deep learning approach. In this work norm-based low complexity joint transmit and receive antenna selection along with its implementation through convolution neural network has been proposed. The proposed supervised deep learning approach to select a group of antennas to enhance capacity and minimize delay has shown encouraging results in MIMO VANET scenario as compared to the conventional antenna selection scheme. [ABSTRACT FROM AUTHOR]

[Display omitted] • A convolution operation in CNNs enable networks to derive significant features using spatial and channel-wise information. A wide range of prior research has canvassed the spatial element to strengthen the representational power of CNN. • In this paper, instead of spatial element we centered our research on channel-wise element. The UNet (38) is state-of-the-art medical image segmentation network which has encoder-decoder structure with number of covolutional layers. For segmentation of tiny structures like nodules, UNet lacks in representation of significant features. The UNet architecture has long skip connections between encoder and decoder at various levels in order to retain the spatial information. • To strengthen the representation power of UNet, we induced the short skip connections called residual connections in the encoder, thereby forming a residual variant of Unet. However, these modifications were not sufficient for effective nodule segmentation. Therefore, we introduce a further modified and improved version of UNet which is termed as SquExUNet. It is a novel nodule segmentation model that utilizes the effectiveness of squeeze and excitation blocks to extract fine-grained information to segment the nodules. • The segmentation framework proposed in this paper utilizes SquExUNet and segments the nodule effectively, but the segmentation framework does emphasize on detect-ability of nodules. Consequently, there arises a need of classification network that can classify the nodule candidates yielded in segmentation framework. • The nodules and non nodule structures are difficult to distinguish in 2D image due to their similar resemblance. Taking that into consideration, we presented an unique approach for detection of nodules using 2D-3D cascaded CNNs. In this detection framework, primarily segmentation framework is used for collection of nodule candidates; and thereafter, a novel 3D classification model termed as 3D-NodNet classifies nodules and non-nodule volumetric cubes. • A novel substantial amount of Indian Lung CT Image Database (ILCID) clinical dataset is included in the study along with publicly available datasets. The early detection of lung cancer is attained with the detection of initial stage nodules (3 − 30 mm) which can exorbitantly increase the 5 -year survival rate of lung cancer patients. Nodules are very small size circumscribed structures in the lungs and are difficult to detect due to their size. The identification of nodule is also more challenging due to similarly resembling structures like non-nodules that contains features which could make it identifiable as a nodule. Therefore, to deal with these challenging issues, we proposed a novel approach for segmentation, classification and detection of lung nodules from CT scan images. In our proposed method we involve a maximum intensity projection technique as a part of image preprocessing method. We demonstrated our experimentation and proposed SquExUNet segmentation model and 3D-NodNet classification model on publicly available Lung Image Database Consortium – Image Database Resource Initiative (LIDC), LNDb Challenge Dataset and purely independent Indian Lung CT Image Database (ILCID) clinical dataset. We proposed a 2D-3D cascaded CNN strategy for detection of nodule that yields the accurately segmented and classified nodule. Results obtained with proposed method indicates that we have successfully detected and segmented the lung nodules effectively, compared to existing lung nodule detection and segmentation algorithms. We achieved a Dice-Coefficient metrics of 0.80 for segmentation of nodule and 90.01 % Sensitivity for nodule detection. [ABSTRACT FROM AUTHOR]

• A systematic review of CNN-based SHM literature is presented in this paper. • The review is categorized into multiple classes depending on the specific SHM. • The existing CNN-based solutions and best practices are illustrated. • Challenges and future research directions of CNN-based methods are presented. With recent advances in non-contact sensing technology such as cameras, unmanned aerial and ground vehicles, the structural health monitoring (SHM) community has witnessed a prominent growth in deep learning-based condition assessment techniques of structural systems. These deep learning methods rely primarily on convolutional neural networks (CNNs). The CNN networks are trained using a large number of datasets for various types of damage and anomaly detection and post-disaster reconnaissance. The trained networks are then utilized to analyze newer data to detect the type and severity of the damage, enhancing the capabilities of non-contact sensors in developing autonomous SHM systems. In recent years, a broad range of CNN architectures has been developed by researchers to accommodate the extent of lighting and weather conditions, the quality of images, the amount of background and foreground noise, and multiclass damage in the structures. This paper presents a detailed literature review of existing CNN-based techniques in the context of infrastructure monitoring and maintenance. The review is categorized into multiple classes depending on the specific application and development of CNNs applied to data obtained from a wide range of structures. The challenges and limitations of the existing literature are discussed in detail at the end, followed by a brief conclusion on potential future research directions of CNN in structural condition assessment. [ABSTRACT FROM AUTHOR]

Heart disease is one of the top ten threats to global health in 2019 according to the WHO. Continuous monitoring of ECG on wearable devices can detect abnormality in the user's heartbeat early, thereby significantly increasing the chance of early intervention which is known to be the key to saving lives. In this paper, we present a set of inter-patient ECG classification methods that use convolutional (CNNs) and spiking neural networks (SNNs). We focused on inter-patient heartbeat classification, in which the model is trained over several patients and then used to infer that for patients not used in training. Raw heartbeat data is used in this paper because most wearable devices cannot deal with complex data preprocessing. A two-steps convolutional neural network testing method is proposed for saving power. For even greater energy-saving, a spiking neural network is also proposed. The latter is obtained from converting the trained CNN model with a less than one percent accuracy drop. The average power of a two-classes SNN is 0.077 W, or 0. 0074 × that of previously proposed neural network-based solutions. [ABSTRACT FROM AUTHOR]

• This research aims to emphasize the huge impact of deep learning models in brain stroke detection and lesion segmentation. • The survey analyses 113 research papers published in different academic research databases in the past few years. • The emerging trends and breakthroughs in stroke detection have been detailed in this evaluation. • It concludes by examining the technical and non-technical challenges faced by researchers and the future implications in stroke detection. In recent years, deep learning algorithms have created a massive impact on addressing research challenges in different domains. The medical field also greatly benefits from the use of improving deep learning models which save time and produce accurate results. This research aims to emphasize the impact of deep learning models in brain stroke detection and lesion segmentation. This is achieved by discussing the state of the art approaches proposed by the recent works in this field. In this study, the advancements in stroke lesion detection and segmentation were focused. The survey analyses 113 research papers published in different academic research databases. The research articles have been filtered out based on specific criteria to obtain the most prominent insights related to stroke lesion detection and segmentation. The features of the stroke lesion vary based on the type of imaging modality. To develop an effective method for stroke lesion detection, the features need to be carefully extracted from the input images. This review takes an attempt to categorize and discuss the different deep architectures employed for stroke lesion detection and segmentation, based on the underlying imaging modality. This further assists in understanding the relevance of the two-deep neural network components in medical image analysis namely Convolutional Neural Network (CNN) and Fully Convolutional Network (FCN). It hints at other possible deep architectures that can be proposed for better results towards stroke lesion detection. Also, the emerging trends and breakthroughs in stroke detection have been detailed in this evaluation. This work concludes by examining the technical and non-technical challenges faced by researchers and indicate the future implications in stroke detection. It could support the bio-medical researchers to propose better solutions for stroke lesion detection. [ABSTRACT FROM AUTHOR]

• An image dataset CPAF of 73635 insect images is established of 20 common insects. • An effective deep neural network model CPAFNet was built to identify the pests. • For getting the best model the key parameters were turned gradually. In order to achieve rapid recognition of the common pests in agriculture and forestry, a new method based on deep convolution neural network was proposed. In this paper, the images of 19 insects and 1 larvae were collected. The data were enhanced by image processing methods such as flipping, rotating, scaling, adding Gussian noise and fancy PCA. The constructed image dataset CPAF had 73,635 insect images, including 4909 original images and 68,726 enhanced images. A 3-folds validation method was used to recognize pest images with VggA, VGG16, Inception V3, ResNet50, as well as CPAFNet, an deep neural network model we designed in this paper, on the CPAF dataset. For the better optimization of identification results, the balanced accuracy was computed and analyzed in this paper. We first compared performance of different models on CPAF dataset. After the same number of iterations of training, the recognition accuracy of the CPAFNet model reached 92.26% when the learning rate was set to 0.02, which is the best performance of all the models participating in the test. The least time spent on training is also the CPAFNet model. Then, the influence of different number of convolution kernels on recognition rate of CPAFNet was analyzed. Results shown that the balance accuracy achieved to 92.63% when the number of convolution kernels corresponding to the convolutional layer group was set to 64-128-256-256. Finally, the influence of different optimization algorithms and dropout probability on training was assessed. Results shown that when the RMSProp algorithm was used and dropout probability was set to 0.8 of CPAFNet, the balance accuracy achieved to 92.63%. In addition, different enhancement algorithms were assessed on pests image recognition of CPAFNet. Results shown that the balance accuracy was decreased from 0.9263 to 0.9152, 0.9125 and 0.9230 on CPAF dataset without expanded data obtained by flipping, rotating and scaling algorithm respectively, which indicate the enhancement algorithms of them can improve the identification precision of pest image. At the same time, the class activation map algorithm was used for feature visualization. Results shown that the CPRAFNet is good for capturing features on pest of CPAF dataset. The results of the model optimization research and the CPAFNet depth model proposed for the CPAF dataset have a good practical significance for the intelligent identification of agricultural and forestry pests. [ABSTRACT FROM AUTHOR]

Availability of accurate land cover information over large areas is essential to the global environment sustainability; digital classification using medium-resolution remote sensing data would provide an effective method to generate the required land cover information. However, low accuracy of existing per-pixel based classification methods for medium-resolution data is a fundamental limiting factor. While convolutional neural networks (CNNs) with deep layers have achieved unprecedented improvements in object recognition applications that rely on fine image structures, they cannot be applied directly to medium-resolution data due to lack of such fine structures. In this paper, considering the spatial relation of a pixel to its neighborhood, we propose a new deep patch-based CNN system tailored for medium-resolution remote sensing data. The system is designed by incorporating distinctive characteristics of medium-resolution data; in particular, the system computes patch-based samples from multidimensional top of atmosphere reflectance data. With a test site from the Florida Everglades area (with a size of 771 square kilometers), the proposed new system has outperformed pixel-based neural network, pixel-based CNN and patch-based neural network by 24.36%, 24.23% and 11.52%, respectively, in overall classification accuracy. By combining the proposed deep CNN and the huge collection of medium-resolution remote sensing data, we believe that much more accurate land cover datasets can be produced over large areas. [ABSTRACT FROM AUTHOR]

This paper presents a system for human posture recognition using a camera and two infrared light sources. It uses as input the combination of the body silhouette and its (invisible to the eye) cast shadows. Conventional video-surveillance methods based on a single camera can fail to infer the correct posture since different postures can look similar under perspective projection. Fortunately, cast body shadows, generated by infrared lights, offer additional posture information that cannot be directly captured by a single camera. Each shadow can be projected on different surfaces (e.g. floor, walls, and furniture) generating complex body projections that represent various shapes within the same posture class. These shadow images are very challenging and difficult to describe with traditional handcrafted features that need to be somewhat invariant to these within-class changes. However, a deep convolution neural network (CNN) is able to learn a better data representation from a large-scale dataset. In the absence of a big real dataset, we propose to use synthetic data for training the CNN classifier. Learning from synthetic data is a challenging task due to the gap between synthetic and real feature distributions. Thus, we propose a normalization technique to bridge the gap and help the classifier to better generalize with real data. We evaluated the proposed system on a new real dataset captured in our laboratory and a simulated dataset generated with computer graphics tools. Experimental results validated the efficiency of the CNN model against other conventional methods. Furthermore, the combination of cast shadows and body silhouette had better performance than using only the body silhouette as expected. [ABSTRACT FROM AUTHOR]

Scene character recognition remains a challenging task due to many interference factors. Considering that characters are composed of a series of parts arranged in certain structures, in this paper, we propose a novel representation termed multi-order co-occurrence activations (MCA) encoded with Fisher Vector (FV), namely MCA-FV. It implicitly models the co-occurrence information of discriminative character parts at different orders to boost the recognition performance. We first extract convolutional activations as local descriptors of character parts from convolutional neural networks (CNNs). Then, we introduce MCA features to capture the multi-order co-occurrence cues among different discriminative character parts. Finally, we apply FV to encode co-occurrence features of each order and obtain a global representation of MCA-FV. The proposed method is evaluated on four scene character datasets including English and Chinese datasets. Experiment results demonstrate the effectiveness of the proposed method for scene character recognition. [ABSTRACT FROM AUTHOR]

Water and energy are essential for human existence, and its rational use should be encouraged. According to the literature review, water consumption directly affects energy consumption and are inseparably linked resources. The energy to water part of the water/energy nexus, increasingly highlighted as an important issue for future planning and strategic policy considerations. Joint consideration of both water/energy domains can identify new options for increasing overall resource use efficiency. This work is part of the project ENERWAT that has as goal to measure in situ the water/energy consumption related with water supply end use in rural and urban dwellings in order to validate the data collected by survey. A methodology for low cost measure and store water/energy consumes was developed. Water, Gas and electricity data was stored in image format. In this paper, a CNN architecture was applied and trained to read water/energy. The models suited their proposed. The achieved accuracy for test set was: water - dozen: 0.98, unit: 0.92; gas: dozen: 0.94, unit: 0.99; and electricity - dozen: 0.99, units 0.99. The more challenge digit was water unit digit due to partial occlusion. It is presented a day of readings and discussed some events. [ABSTRACT FROM AUTHOR]

This paper introduces a multi-view recurrent neural network (MV-RNN) approach for 3D mesh segmentation. Our architecture combines the convolutional neural networks (CNN) and a two-layer long short term memory (LSTM) to yield coherent segmentation of 3D shapes. The imaged-based CNN are useful for effectively generating the edge probability feature map while the LSTM correlates these edge maps across different views and output a well-defined per-view edge image. Evaluations on the Princeton Segmentation Benchmark dataset show that our framework significantly outperforms other state-of-the-art methods. [ABSTRACT FROM AUTHOR]

Object proposal generation is an important step in object detection, obtaining high-quality proposals can effectively improve the performance of detection. In this paper, we propose a semantic, class-specific approach to re-rank object proposals, which can consistently improve the recall performance even with fewer proposals. Specifically, we first extract features for each proposal including semantic segmentation, stereo information, contextual information, CNN-based objectness and low-level cue, and then score them using class-specific weights learned by Structured SVM. The advantages of the proposed model are two-fold: 1) it can be easily merged to existing generators with few computational costs, and 2) it can achieve high recall rate under strict critical even using fewer proposals. Experimental evaluation on the KITTI benchmark demonstrates that our approach significantly improves existing popular generators on recall performance. Moreover, in the experiment conducted for object detection, even with 1500 proposals, our approach can still have higher average precision (AP) than baselines with 5000 proposals. [ABSTRACT FROM AUTHOR]

In the field of defect detection, image processing algorithms and feature extraction algorithms have some limitations, owing to their necessity for extracting a large number of different features of diverse products images. Meanwhile, the images of defective products are less and various. Aiming at these problems, we presented a One-Class classifier based on deep convolution neural network to detect the defect images in this paper. We design a loss function with the penalty term based on Euclidean distance to train the deep convolution neural network model. A hypersphere is used as classification decision surface after setting an appropriate hypersphere radius according to the inspection accuracy. It maps the non-defective products into a hypersphere in a high dimensional feature space, while the defect images are mapped somewhere far from the center of hypersphere. Thus, a One-Class classifier based on convolutional neural network(CNN) model is proposed to detect the defects. Experiments show that the proposed method, with less number of iteration, help build the classifier for image defect detection with high generalization ability and high detection precision. [ABSTRACT FROM AUTHOR]

License Plate Recognition (LPR) is a well-known image processing technology. The goal of this paper is to achieve the same or better accuracy than previous known algorithms while attaining a higher speed of processing and a modular yet simple approach. The engine is mainly designed to work on Lebanese license plates but can easily be trained for others. The results below are obtained by using a deep convolutional neural network cascade for classification (CNN cascade), a CNN with partially connected deep layers for deskewing and a neural network optimized by neuro-evolution for OCR. This resulted in a modular LPR solution that surpasses the conventional solution in terms of speed and accuracy, a deskew module that can straighten double lined plates with far better accuracy than its image processing counterpart and an OCR module that’s optimized for the best speed and accuracy achievable even predicting characters that are unreadable by conventional solutions. On top of that, the whole solution is GPU (Graphical Processing Unit) enabled making it scalable for a large network of cameras with a central processing unit. [ABSTRACT FROM AUTHOR]

Cytology is one of the decisive factors for the early detection of cancer. It is necessary to have a certain number of years of experience to be able to screen tumor cells for cytodiagnosis. However, this diagnosis has poor objectivity because there are so many parts that the judge’s experience and skill is responsible for. In this paper, we present a new method to extract cell nuclei from HE stained images generally used cell staining for Creation of digitized objective indicators in cytology. Our method extracts cell nuclei using combining the input image and the image previously prepared by SVM. Features used in SVM are automatically generated from CNN. Results are demonstrated by experiments using the real images and its ground truths. [ABSTRACT FROM AUTHOR]

In this paper, we consider the problem of detecting communities in topologically incomplete networks (TIN), which are usually observed from real-world networks and where some edges are missing. Existing approaches to community detection always consider the input network as connected. However, more or less, even nearly all, edges are missing in real-world applications, e.g. the protein–protein interaction networks. Clearly, it is a big challenge to effectively detect communities in these observed TIN. At first, we bring forward a simple but useful method to address the problem. Then, we design a structured deep convolutional neural network (CNN) model to better detect communities in TIN. By gradually removing edges of the real-world networks, we show the effectiveness and robustness of our structured deep model on a variety of real-world networks. Moreover, we find that the appropriate choice of hop counts can improve the performance of our deep model in some degree. Finally, experimental results conducted on synthetic data sets also show the good performance of our proposed deep CNN model. [ABSTRACT FROM AUTHOR]

Gender classification (GC) has achieved high accuracy in different experimental evaluations based mostly on inner facial details. However, these results do not generalize well in unrestricted datasets and particularly in cross-database experiments, where the performance drops drastically. In this paper, we analyze the state-of-the-art GC accuracy on three large datasets: MORPH, LFW and GROUPS. We discuss their respective difficulties and bias, concluding that the most challenging and wildest complexity is present in GROUPS. This dataset covers hard conditions such as low resolution imagery and cluttered background. Firstly, we analyze in depth the performance of different descriptors extracted from the face and its local context on this dataset. Selecting the bests and studying their most suitable combination allows us to design a solution that beats any previously published results for GROUPS with the Dago's protocol, reaching an accuracy over 94.2 % , reducing the gap with other simpler datasets. The chosen solution based on local descriptors is later evaluated in a cross-database scenario with the three mentioned datasets, and full dataset 5-fold cross validation. The achieved results are compared with a Convolutional Neural Network approach, achieving rather similar marks. Finally, a solution is proposed combining both focuses, exhibiting great complementarity, boosting GC performance to beat previously published results in GC both cross-database, and full in-database evaluations. [ABSTRACT FROM AUTHOR]

Ship roll motion in high sea states has large amplitudes and nonlinear dynamics, and its prediction is significant for operability, safety, and survivability. This paper presents a novel data-driven methodology to provide a multi-step prediction of ship roll motions in high sea states. A hybrid neural network is proposed that combines long short-term memory (LSTM) and convolutional neural network (CNN) in parallel. The motivation is to extract the nonlinear dynamic characteristics and the hydrodynamic memory information through the advantage of CNN and LSTM, respectively. For the feature selection, the time histories of motion states and wave heights are selected to involve sufficient information. Taken a scaled KCS as the study object, the ship motions in sea state 7 irregular long-crested waves are simulated and used for the validation. The results show that at least one period of roll motion can be accurately predicted. Compared with the single LSTM and CNN methods, the proposed method has better performance in predicting the amplitude of roll angles. Besides, the comparison results also demonstrate that selecting motion states and wave heights as feature space improves the prediction accuracy, verifying the effectiveness of the proposed method. • A data-driven methodology is proposed to achieve the multi-step prediction of ship roll motion in high sea states. • LSTM and CNN are performed in parallel to extract time-dependent and spatio-temporal information from multidimensional inputs. • The CFD method is utilized to generate the motion data in sea state 7 irregular long-peaked waves with different wave directions. • The superiority of selecting both motion states and wave heights as the feature space is demonstated. [ABSTRACT FROM AUTHOR]

Natural gas is vital in the world's energy portfolio and is widely applied to power generation, urban heating, and manufacturing. Forecasting natural gas consumption with high accuracy is thus crucial in order to maintain a reliable supply for various applications. The demand for natural gas often exhibits different seasonal patterns regarding customers of different characteristics. The precision of forecasters will be vulnerably affected without carefully exploring the periodicity of usage. This paper proposes a novel method, Dual Convolution with Seasonal Decomposition Network, for natural gas consumption forecasting. The proposed method applies multiple seasonal-trend decomposition to separate time series into periodic patterns and residual components. In addition, local and global convolution are combined to predict series with significant fluctuations and poor periodicity. Simulations show that on city-level forecasting, the proposed method outperforms state-of-the-art methods in terms of overall prediction accuracy and variation sensitivity regardless of different time intervals. The performance of the method is robust to the forecasting horizon. The method can be deployed in practical circumstances to forecast the natural gas consumption of residential quarters, cities, or even countries in different time spans. • The multiple seasonal-trend decomposition is crucial for extracting regular patterns. • A global and local convolution network is effective in exploring long-short term features. • The autoregressive component improves the input sensitivity of neural networks. • The proposed method outperforms the state-of-the-art on actual data. [ABSTRACT FROM AUTHOR]

Waste management has become a significant issue in most developing countries. Municipal solid waste generation has been steadily increasing over the past decade. Recycling is gaining importance because it is the only method to maintain a healthy and sustainable environment. Furthermore, recycling is not a fully autonomous operation; a significant amount of waste must be performed manually. New and innovative procedures must be implemented to deal with the increasing volume of waste products at recycling facilities. It is suggested that effective solid waste management systems have a practical approach to detecting and classifying waste materials. This study presents CNN and Graph-LSTM, two deep-learning techniques that can recognize typical waste products when handled on a belt conveyor in waste collection systems. This convolutional neural network-based solution is trained to use six object classes: cardboard, metal, glass, plastic, paper, and organic waste. The major advantage is the ability to model long-term dependencies, Improved performance, better generalization, and easy access The experimental findings show that the suggested system can achieve 97.5% accuracy in real-world situations, outperforming existing methods identified in the literature. • New methods are needed to handle growing waste at recycling facilities. • Study introduces CNN & Graph-LSTM for recognizing waste on collection system conveyor belts. • Proposal: Train deep learning model to predict city solid waste. • Deep learning model of CNN & GLSTM predicts solid waste. • Proposed VT-MLH-CNN model has the potential for integration in real-world sorting facilities. [ABSTRACT FROM AUTHOR]

• A new CS-MRI deep unfolded network is proposed to accelerate MR image reconstruction by unfolding FISTA as a deep iterative network. • In order to break the bottleneck of network information transmission, a multi-channel information fusion strategy is used to transmit multi-channel features instead of single-channel images at the input and output of each stage. • A novel channel attention mechanism is proposed to further improve the unfold deep network performance, which can suppress trivial channels by capturing dependencies on channels. • The experimental results show that the proposed method is capable of reconstructing more accurate MR image details with less differential error. Compressed sensing (CS) is often used to accelerate magnetic resonance image (MRI) reconstruction from undersampled k-space data. A novelty deeply unfolded networks (DUNs) based method, designed by unfolding a traditional CS-MRI optimization algorithm into deep networks, can provide significantly faster reconstruction speeds than traditional CS-MRI methods while improving image quality. In this paper, we propose a High-Throughput Fast Iterative Shrinkage Thresholding Network (HFIST-Net) for reconstructing MR images from sparse measurements by combining traditional model-based CS techniques and data-driven deep learning methods. Specifically, the conventional Fast Iterative Shrinkage Thresholding Algorithm (FISTA) method is expanded as a deep network. To break the bottleneck of information transmission, a multi-channel fusion mechanism is proposed to improve the efficiency of information transmission between adjacent network stages. Moreover, a simple yet efficient channel attention block, called Gaussian context transformer (GCT), is proposed to improve the characterization capabilities of deep Convolutional Neural Network (CNN,) which utilizes Gaussian functions that satisfy preset relationships to achieve context feature excitation. T1 and T2 brain MR images from the FastMRI dataset are used to validate the performance of the proposed HFIST-Net. The qualitative and quantitative results showed that our method is superior to those compared state-of-the-art unfolded deep learning networks. The proposed HFIST-Net is capable of reconstructing more accurate MR image details from highly undersampled k-space data while maintaining fast computational speed. [ABSTRACT FROM AUTHOR]

This paper presents a frequency-velocity convolutional neural network (CNN) for rapid, non-invasive 2D shear wave velocity (V S) imaging of near-surface geo-materials. Operating in the frequency-velocity domain allows for significant flexibility in the linear-array, active-source experimental testing configurations used for generating the CNN input, which are normalized dispersion images. While normalized dispersion images retain the most important aspects of near-surface wavefields, they are relatively insensitive to the exact experimental testing configuration used to generate and record the wavefields, accommodating various source types, source offsets, numbers of receivers, and receiver spacings. We demonstrate the effectiveness of the frequency-velocity CNN by applying it to a common near-surface geophysics problem, namely, imaging a two-layer, undulating, soil-over-bedrock interface. The frequency-velocity CNN was trained and tested using 100,000 synthetic near-surface models with variable soil-over-bedrock conditions. Then, the ability of the frequency-velocity CNN to generalize across various acquisition configurations was rigorously tested using thousands of synthetic near-surface models with different acquisition configurations from that of the training set. Lastly, it was applied to experimental field data collected at the Hornsby Bend site in Austin, Texas, USA and found to produce a subsurface 2D image that was in great agreement with ground truth from invasive site characterization data. [ABSTRACT FROM AUTHOR]

Blast furnace (BF) burden surface modeling is the basis for automating precise charging operations of BFs, and it can also be used to predict gas flow distributions based on a burden profile. In this paper, first, a mechanism model is established according to the charging operation, and it is convenient for predicting the burden profile after the charging operation. Then, the Gaussian process regression (GPR) algorithm is used to fuse the charging mechanism model and the radar detection data to better reconstruct the burden profile. Finally, the traditional shape of a burden surface is researched based on the point cloud data of a phased array radar, and 4 classes of burden surfaces are defined and reconstructed. The reconstructed burden surface is classified by expert-defined features and deep features extracted by convolutional neural networks (CNNs). • The two-model drived algorithm based on GPR for 3D burden surface reconstruction. • The 3D burden surface feature is expert-defined and CNN fusion features. • The burden surface classification is through GWO-KELM. [ABSTRACT FROM AUTHOR]

The recognition of learners' speech intention in the online collaborative learning scene is of great significance for exploring the rules of knowledge construction such as knowledge development and emotional communication in the collaborative process. The essence of speech intention recognition is text classification. At present, text classification is mostly based on deep learning model. However, online collaborative discussion has the characteristics of strong contextual semantic relationship and key characteristic words. When only the deep learning method is used for text classification, there may be insufficient acquisition of contextual semantic relations and neglect of key feature words, resulting in a decline in the accuracy of classification results. Therefore, this paper proposes a multi-feature fusion model, which uses BERT to represent the text as a word vector, BiLSTM to extract the context features of the text, CNN to extract the local features of the text, and average pooling model to extract the average representation features of the text. The results show that the overall classification accuracy of the multi-feature fusion model is 84.30%. [ABSTRACT FROM AUTHOR]

• Device identification using several machine learning algorithms such as LR, KNN, SVC, DT, RF, GB, AdaBoost, LGBM, XGB, CNN, and LSTM. • Performance evaluation of models using multiple statistical parameters such as accuracy, precision, recall, F1 scores, etc. • Blockchain-based architecture for the smart home environment incorporating blockchain-based structure and smart contracts. • Blockchain-based architecture for two case studies, i.e., data marketplace and access control mechanism. Due to data collection, there is a potential risk concerning security and privacy, so IoT reliability and survivability are of utmost concern. In this paper, we address the concern using two methods. The first method is device identification, which uses an extensive set of machine learning algorithms for identifying IoT devices. The algorithms include Logistic Regression, K- Nearest Neighbour, Support Vector Classifier, Random Forest, Gradient Boosting, AdaBoost, Light Gradient Boosting Machine, Extreme Gradient Boosting Convolution Neural Networks, and Long Short Term Memory are used for device identification. The performance of these models has been evaluated using multiple statistical measures, and weobserve that LSTM outperforms all other baseline models. The second method proposed for ensuring the survivability of theIoT environment is a blockchain-based architecture for smart homes to ensuretransparency and data protection. [Display omitted] [ABSTRACT FROM AUTHOR]

• A cataract detection model proposed using anterior segment eye images rather than fundus image. • This approach can be used to make a portable diagnostic system using mobile and desktops. • The evaluation was made by targeting iris and pupil as our Region Of Interest. • Adaptive thresholding is then applied to extract the distinguishing features. • A fine-tuned Inception-v3 model is constructed to get the final output classification result. Early detection of cataracts plays a vital role in ensuring the prevention of vision loss. This paper aims to propose an algorithm that will act as an assistive measure in the process of automating cataract disease detection. The majority of the existing works are focused on the utilization of either fundus images, slit lamp images, or visible wavelength images captured using a DSLR camera. The novelty of this proposed algorithm is the capability to deliver equally accurate and precise performance using both normally captured visible wavelength images as well as medically captured anterior segment images which can, in turn, prove to be cost-effective as well. The image pre-processing techniques particularly adaptive thresholding hereby employed provides fast and accurate results on the input dataset fed to the CNN model which in turn is a fine-tuned version of Inception-v3. Here the training of the model has been done using visible wavelength images whereas the validation testing has been done using the anterior segment eye images medically obtained from a hospital. The proposed image pre-processing technique along with the model architecture ensures the achievement of a high classification accuracy of about 95%. Since the model deals with the examination of the anterior segment of the image, cases concerning nuclear cataract, cortical cataract, or a hybrid case involving the detection of both the aforementioned cataract types lie under the range of detection of our system. [ABSTRACT FROM AUTHOR]

• Proposed tunable-Q wavelet method to extract approximate and details of signal and remove redundant information. • Improved the robustness of EEG based epilepsy detection using the deep learning method with CNN model. • Proposed method can achieve 97.57% accuracy, 98.90% sensitivity, 2.10% false positive rate and 10.46-second delay in the real-time seizure detection in testing data. • Combined signal processing method and image classification in this experiment. • Implemented an automatic real-time seizure detection implementation. Epilepsy is a chronic disease caused by sudden abnormal discharge of brain neurons, leading to transient brain dysfunctions. This paper proposed an EEG based real-time approach to detect epilepsy seizures using tunable-Q wavelet transform and convolutional neural network (CNN). Statistical moments and spectral band power were used to reveal the time domain and frequency domain features in EEG, and then were converted into imaged-like data fed into CNN. The proposed approach was evaluated using the database CHB-MIT. The proposed algorithm achieved 97.57% in accuracy, 98.90% in sensitivity, 2.13% in false positive rate and 10.46-second delay. In addition, the proposed method is suitable in real-time implementation. The outcomes indicate that the proposed method can applied to real-time seizure detection in clinical applications. [ABSTRACT FROM AUTHOR]

With the rapid development of deep learning, target detection and segmentation in dirty backgrounds have been readily available. Fluorescent magnetic particle inspection (MPI) based on such technology will be a promising alternative for automated crack defect inspection. Most previous studies in MPI have focused only on crack detection. Instead, we frame it as a crack 3D localization problem, since cracks on non-machined surfaces of metal parts need to be polished and re-inspected, which relies on their 3D positions. Although good results were obtained in defect detection, it is still challenging to perform pixel-level segmentation of micro-cracks from the large background to obtain crack 2D pixels for 3D reconstruction. This paper proposes a two-stage convolutional neural network (CNN) method for metal parts crack defect detection and segmentation at the image-pixel level. The first stage detects and crops the potential cracks to a small area, and the second stage can learn the context of cracks in the detected patches. A window-based stereo matching method is then used to find matching pixels of cracks and to map crack image plane points to the 3D world points. We also illustrate the entire system's model deployment and signaling work to apply these methods. Both computational and experimental results based on the system are presented for validation. The training precision of target detection reaches 96.3%, its average precision reaches 85.4%, and the average precision reaches 98.3% when the Intersection-over-Union (IoU) threshold is 0.5. The Dice score reaches 94% in pixel-level segmentation, and the average precision is 99.3% when the probability threshold is set to 0.5. The corresponding efficiency reaches 19 FPS and 18 FPS, and the mean absolute errors of 3D coordinates of reconstructed crack defects are all within 1 mm in X-, Y- and Z- directions. • Crack defect detection and 3D location for fluorescent magnetic particle inspection. • A two-stage convolutional neural network with object detection. • Visual system calibration and efficiency 3D reconstruction. • Experiments and results for detection, segmentation, and 3D reconstruction. • Deep learning model deployment, signal communication, computational efficiency. [ABSTRACT FROM AUTHOR]

• A new semantic segmentation model of parallel network architecture is proposed to segment grape clusters, peduncles and occlusions. • A feature fusion module (FFM) is designed to coordinate the inconsistent characteristics of CNN and Transformer branches. • The demonstrated performance indicated that fusion model has the best segmentation effect on grape images. Semantic segmentation is a kind of classification at the pixel level, which has a remarkable effect in dealing with complex scenes. Most popular semantic segmentation models are based on Convolutional Neural Network (CNN) and Transformer structures. Convolution operations and attention modules are good at extracting local and global features. In this paper, we propose a parallel network structure, termed DualSeg, which leverages the advantages of CNN at local processing and Transformer at global interaction. And a feature fusion module (FFM) is designed to fuse parameters between modules in different branches to retain local and global features to the maximum extent. In the vineyards, the grape clusters and grape peduncles are often heavily obscured by leaves, branches and clusters, making it difficult to distinguish them accurately. Therefore, the scene is taken as an example to verify the segmentation effect of the model. In the experiment, the DualSeg model and mainstream segmentation models are used to test the segmentation effect of different models in this scene. And the experimental results showed that the DualSeg model had the best segmentation effect of all models. Specifically, the model had an IoU value of grape peduncle segmentation of 72.1%, which was more than 3.9% higher than that of other models in the case of 80 K iterations. In addition, the mIoU value of this model is 83.7%, which was the highest value compared with other models. The demonstrated performance of DualSeg model indicated that harvesting robots could be use it. [ABSTRACT FROM AUTHOR]

X-ray Computed Tomography (CT) is a commonly used imaging technique for non-destructive inspection of manufactured objects. However, a full CT scan requires a long acquisition time, making this method unsuitable for inline applications. In contrast to X-ray CT, inspection can be performed directly in the projection space, using simulated X-ray projections of a reference model of the manufactured object. However, to effectively compare simulated and measured projections, an accurate 3D pose estimation of the object and consequent alignment between the measured object and the reference model are crucial. In this paper, we present a fast method to estimate the 3D pose of a measured object based on convolutional neural networks (CNNs). Through experiments on synthetic and measured data, we demonstrate that our method allows estimating the 3D pose of the object with sub-pixel accuracy. Even if very few projections are available, our approach is comparable to CT-based methods for registration, and outperforms state-of-the-art deep learning methods for radiograph-based pose estimation. • Our framework allows projection-based 3D inspection from very few images. • CNN-based 3D pose estimation from one or multiple radiographs. • High accuracy and precision demonstrated on simulated and measured data. • Comparable to CT-based registration. • Outperforms state-of-the-art DL-based 3D pose estimation from radiographs. [ABSTRACT FROM AUTHOR]

• An end-to-end method based on convolutional neural network and transformer (ECT-net) is proposed to detect ECG characteristic waveform under the inter-patient pattern. • Transformer is applied to detect ECG characteristic waveform for the first time. • The proposed method shows excellent performance in different pathological ECG datasets. The electrocardiogram (ECG) characteristic waveforms include P wave, QRS wave, and T wave. The detection of ECG characteristic waveforms is particularly important in automatic heart disease diagnosis. Although researchers have proposed many detection methods, the detection under the inter-patient pattern is still a challenge. This paper proposed an end-to-end ECG waveform detection method (ECT-net) based on convolutional neural network (CNN) and transformer. We constructed the feature extractor by a group of convolutional layers to extract the local information of ECG signals and generate the input embedding of the transformer encoder. Meanwhile, we adopted the transformer encoder to extract the long-term time-dependent representation between heartbeats, which can make up for the limitation of the CNN in obtaining global information. This structure ensures that both spatial and temporal features from ECG signals are achieved. For precise detection, we employed a corresponding decoder and skip structure to integrate low-level features and high-level features. The model was trained and verified on the China Physiological Signal Challenge 2018 database (CPSC-DB). Our method performed best among comparison methods and achieved F1 scores of 94.27% on the P wave, 97.32% on the QRS wave, and 93.92% on the T wave. The ablation result shows that the transformer encoder is beneficial for detecting each characteristic waveform. To assess the generalization ability of the method, it was also evaluated on inter-patient datasets with different types of cardiac diseases, and experimental results demonstrated that the proposed method is effective for all of them. The proposed method applies the transformer encoder to the detection of ECG characteristic waveforms for the first time and achieves competitive performance. [ABSTRACT FROM AUTHOR]

Phase-sensitive optical time domain reflectometer (Φ-OTDR) can detect the occurrence of external vibration, but lacks of ability to classify multiple similar events. With the help of machine learning method, the events can be basically classified. An event recognition method based on deep learning and traditional classifier is proposed in this paper to further improve the classification accuracy. Firstly, the temporal-spatial data matrix from Φ-OTDR is directly sent to convolutional neural network (CNN) and the data features can be obtained automatically. Then these data features are transmitted to a traditional classifier for further classification. The traditional classifier can find a better classification hyperplane to improve the classification accuracy. In this paper, the most suitable classifier is support vector machine (SVM). As CNN works as a black box, T-distributed stochastic neighbor embedding (T-SNE) method and gradient-weighted class activation mapping (Grad-CAM) method are applied to visualize CNN's working process, which illustrates the correctness of feature extraction. Experimental results based on 11,997 images under eight event categories show that the classification accuracy can be improved by the traditional classifier. The CNN+SVM strategy performs the best and reaches 94.17 % accuracy, which offers 2.04 % improvement compared with using CNN alone. [ABSTRACT FROM AUTHOR]

• In this paper, we propose a matching network which builds a connection be- tween template characters and handwritten characters inspired by the human learning process of writing Chinese characters. The matching network replace the parameters in the softmax regression layer with the features extracted from the template character images. After the training process has been finished, the powerful discriminative features help us to generalize the predictive power not just to new data, but to entire new classes that never appear in the training set before. Experiments performed on the ICDAR-2013 offline HCCR datasets have shown that the proposed method achieves a comparable performance to current CNN-based classifiers. Besides, the matching network has a very promising gen- eralization ability to the new classes that never appear in the existing training set. Just like its remarkable achievements in many computer vision tasks, the convolutional neural networks (CNN) provide an end-to-end solution in handwritten Chinese character recognition (HCCR) with great success. However, the process of learning discriminative features for image recognition is difficult in cases where little data is available. In this paper, we propose a matching network which builds a connection between template characters and handwritten characters inspired by the human learning process of writing Chinese characters. The matching network replaces the parameters in the softmax regression layer with the features extracted from the template character images. After the training process has been finished, the powerful discriminative features help us to generalize the predictive power not just to new data, but to entire new Chinese characters that never appear in the training set before. Experiments performed on the ICDAR-2013 offline HCCR datasets have shown that the proposed method achieves a comparable performance to current CNN-based classifiers. Besides, the matching network has a very promising generalization ability to new Chinese characters that never appear in the existing training set. [ABSTRACT FROM AUTHOR]

• Three typical rough handling of parcels in the express industry were summarized. • The method to identify the kind of rough handling with 96% accuracy was proposed. • A hardware device to collect three-axis acceleration of parcels was developed. • A new feature expression was designed to suit the actual physical background. • A CNN model commonly used in visual recognition was modified to use in the method. Rough Handling of Express Parcels (RHEP) increases the risk of courier cargo damage and hurts the reputation of the express industry. Meanwhile, RHEP indirectly caused excessive use of packaging materials and cushioning materials, thereby aggravating environmental and social problems, such as waste disposal pressure and waste of resources. The prerequisite to prevent RHEP is recognition of it. Three typical types of RHEP (dropping, kicking fast and throwing) are discussed in this paper. For these types of RHEP, an intelligent recognition method is proposed. The main idea is to window the acceleration data of the package and extract the features in the window, then arrange these feature data into a three-dimensional matrix as the input of the CNN, and finally obtain the recognition result. In the study, mean, variance, kurtosis, skewness, dynamic range, short-term energy, and zero-crossing rate are considered to be good features in the issues dealt with in this paper. The feature data are organized in a way that different features occupy different channels. Within the uniform channel, three rows correspond to three axes, and the time window order corresponds to consecutive columns. This arrangement of feature data can take full advantage of convolution operations to mine the potential information of acceleration signals with time series characteristics and spatial characteristics. Experimental results show that the recognition accuracy of this method can reach 93.2–96.12% steadily, which is better than the performance of directly arranging features into vectors and sending it to the fully connected network (recognition accuracy is 85–95% and fluctuates greatly). Combining the recognition results of this method with the time and place of RHEP can more clearly analyze the causes of RHEP and propose targeted preventive measures. [ABSTRACT FROM AUTHOR]

In order to make full use of the complementary information of infrared and visible images, this paper proposes an infrared and visible image fusion algorithm based on supervised convolutional neural network (CNN). The network consists of coding layer, fusion layer, decoding layer and output layer. The coding layer contains two convolution layers and two dense modules, which are used to extract the features of the input images. In the fusion layer, an improved algorithm weighted by brightness is used to fuse the extracted feature maps. The decoding layer consists of three convolution layers. In the training process, this paper takes the fusion results of existing literature as the training labels, and improves a loss function based on square loss and structural similarity. The decoded infrared image, decoded visible image and decoded feature image are trained together. The fused image is obtained by weighting these three images. Experimental results show that the proposed method can better preserve both the clear target and detailed information of infrared and visible images. Compared with existing fusion methods, experiments also demonstrate the superiority of the proposed method over the state-of-the-art methods in objective metrics. [ABSTRACT FROM AUTHOR]

To detect new malware, machine learning approaches require many training samples. These training samples contribute to build an accurate model. To maintain the accuracy, collecting comprehensive samples continuously is very important. However, new malicious samples appear one after another, and thereby making it difficult. Hence, actual small training samples do not likely to represent the entire population adequately. Despite this gap between ideal and reality, few studies have addressed this practical problem in macro malware. To enhance small training samples, data augmentation is efficient in the field of image recognition. Data augmentation with Generative Adversarial Networks (GANs) is a reasonable approach for oversampling the minority class. A major difficulty of GANs is to generate fake samples that represent the context. This paper attempts to generate fake text vectors with Paragraph Vector to enhance small training samples. Paragraph Vector is a model to convert text into vectors, which represents the context and numerical distance. These features allow to directly vary each element of the vectors. Our method adds random noise to the vectors to generate fake text vectors which represent the context. This paper applies this technique to detect new malicious VBA (Visual Basic for Applications) macros to address the practical problem. This generic technique could be used for not only malware detection, but also any imbalanced and contextual data. To simulate small training samples, we reduce the malicious samples, and generate fake samples from the reduced ones. The experimental result shows that the fake samples enhance our model, and improve the detection rate. [ABSTRACT FROM AUTHOR]

• A new benchmarking dataset was built from 18157 breast density images, manually segmented into 4 levels based on Breast Imaging and Reporting Data System (BI-RADS): A (fatty), B (fibrous gland), C (uneven dense), and D (dense). • This paper proposes an improved convolutional neural network (CNN) framework that integrates innovative SE-Attention mechanism to learn discriminative features, aiming for automatic breast density classification in mammography. • Transfer learning and auxiliary loss was introduced in the SE-Attention mechanism. Background and objective: Breast density (BD) is an independent predictor of breast cancer risk factor. The automatic classification of BD has yet to resolve. In this paper, we propose an improved convolutional neural network (CNN) framework that integrates innovative SE-Attention mechanism to learn discriminative features, aiming for automatic BD classification in mammography. Methods: A new benchmarking dataset was constructed from 18157 BD images, manually segmented into 4 levels based on Breast Imaging and Reporting Data System (BI-RADS): A (fatty), B (fibro-glandular), C (heterogeneously dense) and D (extremely dense). The proposed method consists of three main phases: (i) data enhancement and normalization of breast images (ii) SE-Attention training for feature re-calibration and fusion to better classify density and (iii) designing the auxiliary loss. We adopt an attention approach where SE-Attention mechanism is used to learn the density features, which is different from previous works. Results: Experimental results demonstrate that the proposed framework obtains higher classification accuracy than the original network, such as Inception-V4, ResNeXt, DenseNet, increasing the performance from 89.97% to 92.17%, 89.64% to 91.57%, 89.20% to 91.79% respectively. Among them, improved Inception-V4 possesses the highest accuracy meanwhile DenseNet improves in the largest extent, both the original and improved methods are more effective than other state-of-the-art image descriptors regarding classification. Conclusions: We insist that our method will help radiologists provide reliable BD diagnostic services at the expert level, allowing them to focus on patients who are really in need. [ABSTRACT FROM AUTHOR]

Multi-view image sensing is currently gaining momentum, fostered by new applications such as autonomous vehicles and self-propelled robots. In this paper, we prototype and evaluate a multi-view smart vision system for object recognition. The system exploits an optimized Multi-View Convolutional Neural Network (MVCNN) in which the processing is distributed among several sensors (heads) and a camera body. The main challenge for designing such a system comes from the computationally expensive workload of real-time MVCNNs which is difficult to support with embedded processing and high frame rates. This paper focuses on the decisions to be taken for distributing an MVCNN on the camera heads, each camera head embedding a Field-Programmable Gate Array (FPGA) for processing images on the stream. In particular, we show that the first layer of the AlexNet network can be processed at the nearest of the sensors, by performing a Direct Hardware Mapping (DHM) using a dataflow model of computation. The feature maps produced by the first layers are merged and processed by a camera central processing node that executes the remaining layers. The proposed system exploits state-of-the-art deep learning optimization methods, such as parameter removing and data quantization. We demonstrate that accuracy drops caused by these optimizations can be compensated by the multi-view nature of the captured information. Experimental results conducted with the AlexNet CNN show that the proposed partitioning and resulting optimizations can fit the first layer of the multi-view network in low-end FPGAs. Among all the tested configurations, we propose 2 setups with an equivalent accuracy compared to the original network on the ModelNet40 dataset. The first one is composed of 4 cameras based on a Cyclone III E120 FPGA to embed the least expensive version in terms of logic resources while the second version requires 2 cameras based on a Cyclone 10 GX220 FPGA. This distributed computing with workload reduction is demonstrated to be a practical solution when building a real-time multi-view smart camera processing several frames per second. [ABSTRACT FROM AUTHOR]

Financial time series forecasting is undoubtedly the top choice of computational intelligence for finance researchers in both academia and the finance industry due to its broad implementation areas and substantial impact. Machine Learning (ML) researchers have created various models, and a vast number of studies have been published accordingly. As such, a significant number of surveys exist covering ML studies on financial time series forecasting. Lately, Deep Learning (DL) models have appeared within the field, with results that significantly outperform their traditional ML counterparts. Even though there is a growing interest in developing models for financial time series forecasting, there is a lack of review papers that solely focus on DL for finance. Hence, the motivation of this paper is to provide a comprehensive literature review of DL studies on financial time series forecasting implementation. We not only categorized the studies according to their intended forecasting implementation areas, such as index, forex, and commodity forecasting, but we also grouped them based on their DL model choices, such as Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), and Long-Short Term Memory (LSTM). We also tried to envision the future of the field by highlighting its possible setbacks and opportunities for the benefit of interested researchers. • We reviewed all searchable articles of deep learning (DL) for financial time series forecasting. • RNN based DL models (LSTM and GRU included) are the most common. • We compared DL models according to their performances in different forecasted asset classes. • To best of our knowledge, this is the first comprehensive DL survey for financial time series forecasting. • We provided current status of DL in financial time series forecasting, also highlighted the future opportunities. [ABSTRACT FROM AUTHOR]