Search

Deep learning, a crucial technique for achieving artificial intelligence (AI), has been successfully applied in many fields. The gradual application of the latest architectures of deep learning in the field of time series forecasting (TSF), such as Transformers, has shown excellent performance and results compared to traditional statistical methods. These applications are widely present in academia and in our daily lives, covering many areas including forecasting electricity consumption in power systems, meteorological rainfall, traffic flow, quantitative trading, risk control in finance, sales operations and price predictions for commercial companies, and pandemic prediction in the medical field. Deep learning-based TSF tasks stand out as one of the most valuable AI scenarios for research, playing an important role in explaining complex real-world phenomena. However, deep learning models still face challenges: they need to deal with the challenge of large-scale data in the information age, achieve longer forecasting ranges, reduce excessively high computational complexity, etc. Therefore, novel methods and more effective solutions are essential. In this paper, we review the latest developments in deep learning for TSF. We begin by introducing the recent development trends in the field of TSF and then propose a new taxonomy from the perspective of deep neural network models, comprehensively covering articles published over the past five years. We also organize commonly used experimental evaluation metrics and datasets. Finally, we point out current issues with the existing solutions and suggest promising future directions in the field of deep learning combined with TSF. This paper is the most comprehensive review related to TSF in recent years and will provide a detailed index for researchers in this field and those who are just starting out. [ABSTRACT FROM AUTHOR]

Featured Application: The Deep Neural Network model employed in this work will help build a system for the pre-diagnosis of retinopathies that can lead to blindness. The main intention of this type of system is to support the work performed by ophthalmology specialists. Computer-based pre-diagnosis of diseases through medical imaging is a task worked on for many years. The so-called fundus images stand out since they do not have uniform illumination and are highly sensitive to noise. One of the diseases that can be pre-diagnosed through fundus images is age-related macular degeneration, which initially manifests as the appearance of lesions called drusen. Several ways of pre-diagnosing macular degeneration have been proposed, methods based entirely on the segmentation of drusen with prior image processing have been designed and applied, and methods based on image pre-processing and subsequent conversion to feature vectors, or patterns, to be classified by a Machine-Learning model have also been developed. Finally, in recent years, the use of Deep-Learning models, particularly Convolutional Networks, has been proposed and used in classification problems where the data are only images. The latter has allowed the so-called transfer learning, which consists of using the learning achieved in the solution of one problem to solve another. In this paper, we propose the use of transfer learning through the Xception Deep Convolutional Neural Network to detect age-related macular degeneration in fundus images. The performance of the Xception model was compared against six other state-of-the-art models with a dataset created from images available in public and private datasets, which were divided into training/validation and test; with the training/validation set, the training was made using 10-fold cross-validation. The results show that the Xception neural network obtained a validation accuracy that surpasses other models, such as the VGG-16 or VGG-19 networks, and had an accuracy higher than 80% in the test set. We consider that the contributions of this work include the use of a Convolutional Neural Network model for the detection of age-related macular degeneration through the classification of fundus images in those affected by AMD (drusen) and the images of healthy patients. The performance of this model is compared against other methods featured in the state-of-the-art approaches, and the best model is tested on a test set outside the training and validation set. [ABSTRACT FROM AUTHOR]

Image restoration under adverse weather conditions poses a challenging task. Previous research efforts have predominantly focused on eliminating rain and fog phenomena from images. However, snow, being another common atmospheric occurrence, also significantly impacts advanced computer vision tasks such as object detection and semantic segmentation. Recently, there has been a surge of methods specifically targeting snow removal, with the majority employing visual Transformers as the backbone network to enhance restoration effectiveness. Nevertheless, due to the quadratic computations required by Transformers to model long-range dependencies, this significantly escalates the time and space consumption of deep learning models. To address this issue, this paper proposes an efficient snow removal Transformer with a global windowing network (SGNet). This method forgoes the localized windowing strategy of previous visual Transformers, opting instead to partition the image into multiple low-resolution subimages containing global information using wavelet sampling, thereby ensuring higher performance while reducing computational overhead. Extensive experimentation demonstrates that our approach achieves outstanding performance across a wide range of benchmark datasets and can rival methods employing CNNs in terms of computational cost. [ABSTRACT FROM AUTHOR]

Electroencephalography (EEG) signals can be used as a neuroimaging indicator to analyze brain-related diseases and mental states, such as schizophrenia, which is a common and serious mental disorder. However, the main limiting factor of using EEG data to support clinical schizophrenia diagnosis lies in the inadequacy of both objective characteristics and effective data analysis techniques. Random matrix theory (RMT) and its linear eigenvalue statistics (LES) can provide an effective mathematical modeling method for exploring the statistical properties of non-stationary nonlinear systems, such as EEG signals. To obtain an accurate classification and diagnosis of schizophrenia, this paper proposes a LES-based deep learning network scheme in which a series of random matrixes, consisting of EEG data sliding window sampling and their eigenvalues, are employed as features for deep learning. Due to the fact that the performance of the LES-based scheme is sensitive to the LES's test function, the proposed LES-based deep learning network is embedded with two ways of combining LES's test functions with learning techniques: the first is to have the LES's test function assigned, while, using the second way, the optimal LES's test function should be solved in a functional optimization problem. In this paper, various test functions and different optimal learning methods were coupled in experiments. Our results revealed a binary classification accuracy of nearly 90% in distinguishing between healthy controls (HC) and patients experiencing the first episode of schizophrenia (FES). Additionally, we achieved a ternary classification accuracy of approximately 70% by including clinical high risk for psychosis (CHR). The LES-embedded approach yielded notably higher classification accuracy compared to conventional machine learning methods and standard convolutional neural networks. As the performance of schizophrenia classification is strongly influenced by test functions, a functional optimization problem was proposed to identify an optimized test function, and an approximated parameter optimization problem was introduced to limit the search area of suitable basis functions. Furthermore, the parameterization test function optimization problem and the deep learning network were coupled to be synchronously optimized during the training process. The proposal approach achieved higher classification accuracy rates of 96.87% between HC and FES, with an additional 89.06% accuracy when CHR was included. The experimental studies demonstrated that the proposed LES-based method was significantly effective for schizophrenic EEG data classification. [ABSTRACT FROM AUTHOR]

This paper discusses the challenges of the hyperparameter tuning in deep learning models and proposes a green approach to the neural architecture search process that minimizes its environmental impact. The traditional approach of neural architecture search involves sweeping the entire space of possible architectures, which is computationally expensive and time-consuming. Recently, to address this issue, performance predictors have been proposed to estimate the performance of different architectures, thereby reducing the search space and speeding up the exploration process. The proposed approach aims to develop a performance predictor by training only a small percentage of the possible hyperparameter configurations. The suggested predictor can be queried to find the best configurations without training them on the dataset. Numerical examples of image denoising and classification enable us to evaluate the performance of the proposed approach in terms of performance and time complexity. [ABSTRACT FROM AUTHOR]

Quantization is a widespread method for reducing the inference time of neural networks on mobile Central Processing Units (CPUs). Eight-bit quantized networks demonstrate similarly high quality as full precision models and perfectly fit the hardware architecture with one-byte coefficients and thirty-two-bit dot product accumulators. Lower precision quantizations usually suffer from noticeable quality loss and require specific computational algorithms to outperform eight-bit quantization. In this paper, we propose a novel 4.6-bit quantization scheme that allows for more efficient use of CPU resources. This scheme has more quantization bins than four-bit quantization and is more accurate while preserving the computational efficiency of the later (it runs only 4% slower). Our multiplication uses a combination of 16- and 32-bit accumulators and avoids multiplication depth limitation, which the previous 4-bit multiplication algorithm had. The experiments with different convolutional neural networks on CIFAR-10 and ImageNet datasets show that 4.6-bit quantized networks are 1.5–1.6 times faster than eight-bit networks on the ARMv8 CPU. Regarding the quality, the results of the 4.6-bit quantized network are close to the mean of four-bit and eight-bit networks of the same architecture. Therefore, 4.6-bit quantization may serve as an intermediate solution between fast and inaccurate low-bit network quantizations and accurate but relatively slow eight-bit ones. [ABSTRACT FROM AUTHOR]

Sentiment analysis has increasingly gained significance in commercial settings, driven by the rising impact of reviews on purchase decision-making in recent years. This research conducts a thorough examination of the suitability of machine learning and deep learning approaches for sentiment analysis, using Romanian reviews as a case study, with the aim of gaining insights into their practical utility. A comprehensive, multi-level analysis is performed, covering the document, sentence, and aspect levels. The main contributions of the paper refer to the in-depth exploration of multiple sentiment analysis models at three different textual levels and the subsequent improvements brought with respect to these standard models. Additionally, a balanced dataset of Romanian reviews from twelve product categories is introduced. The results indicate that, at the document level, supervised deep learning techniques yield the best outcomes (specifically, a convolutional neural network model that obtains an AUC value of 0.93 for binary classification and a weighted average F1-score of 0.77 in a multi-class setting with 5 target classes), albeit with increased resource consumption. Favorable results are achieved at the sentence level, as well, despite the heightened complexity of sentiment identification. In this case, the best-performing model is logistic regression, for which a weighted average F1-score of 0.77 is obtained in a multi-class polarity classification task with three classes. Finally, at the aspect level, promising outcomes are observed in both aspect term extraction and aspect category detection tasks, in the form of coherent and easily interpretable word clusters, encouraging further exploration in the context of aspect-based sentiment analysis for the Romanian language. [ABSTRACT FROM AUTHOR]

Applied computing and artificial intelligence methods have been attracting growing interest in recent years due to their effectiveness in solving technical problems. Hao et al. [[14]] present an unsupervised fault diagnosis methodology to leverage the generated MPCs of different working conditions to diagnose the actual unlabeled MPCs. The paper by Ainapure et al. [[17]] proposes a new cross-domain fault diagnosis method with enhanced robustness. The paper by Saeed et al. [[1]] proposes an approach to building an AutoML data-dependent CNN model (DeepPCANet) customized for DR screening automatically. [Extracted from the article]

Images are a convenient way for humans to obtain information and knowledge, but they are often destroyed throughout the collection or distribution process. Therefore, image processing evolves as the need arises, and color image processing is a broad and active field. A color image includes three distinct but closely related channels (red, green, and blue (RGB)). Compared to directly expressing color images as vectors or matrices, the quaternion representation offers an effective alternative. There are several papers and works on this subject, as well as numerous definitions, hypotheses, and methodologies. Our observations indicate that the quaternion representation method is effective, and models and methods based on it have rapidly developed. Hence, the purpose of this paper is to review and categorize past methods, as well as study their efficacy and computational examples. We hope that this research will be helpful to academics interested in quaternion representation. [ABSTRACT FROM AUTHOR]

The classification of certain agricultural species poses a formidable challenge due to their inherent resemblance and the absence of dependable visual discriminators. The accurate identification of these plants holds substantial importance in industries such as cosmetics, pharmaceuticals, and herbal medicine, where the optimization of essential compound yields and product quality is paramount. In response to this challenge, we have devised an automated classification system based on deep learning principles, designed to achieve precision and efficiency in species classification. Our approach leverages a diverse dataset encompassing various cultivars and employs the Parallel Artificial Multiple Intelligence System–Ensemble Deep Learning model (P-AMIS-E). This model integrates ensemble image segmentation techniques, including U-Net and Mask-R-CNN, alongside image augmentation and convolutional neural network (CNN) architectures such as SqueezeNet, ShuffleNetv2 1.0x, MobileNetV3, and InceptionV1. The culmination of these elements results in the P-AMIS-E model, enhanced by an Artificial Multiple Intelligence System (AMIS) for decision fusion, ultimately achieving an impressive accuracy rate of 98.41%. This accuracy notably surpasses the performance of existing methods, such as ResNet-101 and Xception, which attain 93.74% accuracy on the testing dataset. Moreover, when applied to an unseen dataset, the P-AMIS-E model demonstrates a substantial advantage, yielding accuracy rates ranging from 4.45% to 31.16% higher than those of the compared methods. It is worth highlighting that our heterogeneous ensemble approach consistently outperforms both single large models and homogeneous ensemble methods, achieving an average improvement of 13.45%. This paper provides a case study focused on the Centella Asiatica Urban (CAU) cultivar to exemplify the practical application of our approach. By integrating image segmentation, augmentation, and decision fusion, we have significantly enhanced accuracy and efficiency. This research holds theoretical implications for the advancement of deep learning techniques in image classification tasks while also offering practical benefits for industries reliant on precise species identification. [ABSTRACT FROM AUTHOR]

Given an observed low frame rate video, video frame interpolation (VFI) aims to generate a high frame rate video, which has smooth video frames with higher frames per second (FPS). Most existing VFI methods often focus on generating one frame at a specific timestep, e.g., 0.5, between every two frames, thus lacking the flexibility to increase the video's FPS by an arbitrary scale, e.g., 3. To better address this issue, in this paper, we propose an arbitrary timestep video frame interpolation (ATVFI) network with time-dependent decoding. Generally, the proposed ATVFI is an encoder–decoder architecture, where the interpolation timestep is an extra input added to the decoder network; this enables ATVFI to interpolate frames at arbitrary timesteps between input frames and to increase the video's FPS at any given scale. Moreover, we propose a data augmentation method, i.e., multi-width window sampling, where video frames can be split into training samples with multiple window widths, to better leverage training frames for arbitrary timestep interpolation. Extensive experiments were conducted to demonstrate the superiority of our model over existing baseline models on several testing datasets. Specifically, our model trained on the GoPro training set achieved 32.50 on the PSNR metric on the commonly used Vimeo90k testing set. [ABSTRACT FROM AUTHOR]

At present, decision-making solutions developed based on deep learning (DL) models have received extensive attention in predictive maintenance (PM) applications along with the rapid improvement of computing power. Relying on the superior properties of shared weights and spatial pooling, convolutional neural networks (CNNs) can learn effective representations of health states from industrial data. Many developed CNN-based schemes, such as advanced CNNs that introduce residual learning and multi-scale learning, have shown good performance in health states recognition tasks under the assumption that all the classes are known. However, these schemes have no ability to deal with new abnormal samples that belong to state classes not part of the training set. In this paper, a collective decision framework for different CNNs is proposed. It is based on a one-vs.-rest network (OVRN) to simultaneously achieve classification of known and unknown health states. OVRNs learn class-specific discriminative features and enhance the ability to reject new abnormal samples incorporated to different CNNs. According to the validation results on the public dataset of the Tennessee Eastman process (TEP), the proposed CNN-based decision schemes incorporating an OVRN have outstanding recognition ability for samples of unknown heath states while maintaining satisfactory accuracy on known states. The results show that the new DL framework outperforms state-of-the-art CNNs, and the one based on residual and multi-scale learning has the best overall performance. [ABSTRACT FROM AUTHOR]

Hybrid approaches combining machine learning with traditional inverse problem solution methods represent a promising direction for the further development of inverse modeling algorithms. The paper proposes an approach to emission source identification from measurement data for advection–diffusion–reaction models. The approach combines general-type source identification and post-processing refinement: first, emission source identification by measurement data is carried out by a sensitivity operator-based algorithm, and then refinement is done by incorporating a priori information about unknown sources. A general-type distributed emission source identified at the first stage is transformed into a localized source consisting of multiple point-wise sources. The second, refinement stage consists of two steps: point-wise source localization and emission rate estimation. Emission source localization is carried out using deep learning with convolutional neural networks. Training samples are generated using a sensitivity operator obtained at the source identification stage. The algorithm was tested in regional remote sensing emission source identification scenarios for the Lake Baikal region and was able to refine the emission source reconstruction results. Hence, the aggregates used in traditional inverse problem solution algorithms can be successfully applied within machine learning frameworks to produce hybrid algorithms. [ABSTRACT FROM AUTHOR]

Control charts with conditional expected value (CEV) can be used with novel statistical techniques to monitor the means of moderately and lowly censored data. In recent years, machine learning and deep learning have been successfully combined with quality technology to solve many process control problems. This paper proposes a residual control chart combining a convolutional neural network (CNN) and support vector regression (SVR) for type-I censored data with the Weibull model. The CEV and exponentially weighted moving average (EWMA) statistics are used to generate training data for the CNN and SVR. The average run length shows that the proposed chart approach outperforms the traditional EWMA CEV chart approach in various shift sizes and censored rates. The proposed chart approach is suitable to be used in detecting small shift size for highly censored data. An illustrative example presents the application of the proposed method in an electronics industry. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) have gained recognition for their remarkable performance across various tasks. However, the sheer number of parameters and the computational demands pose challenges, particularly on edge devices with limited processing power. In response to these challenges, this paper presents a novel approach aimed at enhancing the efficiency of deep learning models. Our method introduces the concept of accuracy and efficiency coefficients, offering a fine-grained control mechanism to balance the trade-off between network accuracy and computational efficiency. At our core is the Rewarded Meta-Pruning algorithm, guiding neural network training to generate pruned model weight configurations. The selection of this pruned model is based on approximations of the final model's parameters, and it is precisely controlled through a reward function. This reward function empowers us to tailor the optimization process, leading to more effective fine-tuning and improved model performance. Extensive experiments and evaluations underscore the superiority of our proposed method when compared to state-of-the-art techniques. We conducted rigorous pruning experiments on well-established architectures such as ResNet-50, MobileNetV1, and MobileNetV2. The results not only validate the efficacy of our approach but also highlight its potential to significantly advance the field of model compression and deployment on resource-constrained edge devices. [ABSTRACT FROM AUTHOR]

Resistance spot welding poses potential challenges for automotive manufacturing enterprises with regard to ensuring the real-time and accurate quality detection of each welding spot. Nowadays, many machine learning and deep learning methods have been proposed to utilize monitored sensor data to solve these challenges. However, poor detection results or process interpretations are still unaddressed key issues. To bridge the gap, this paper takes the automotive bodies as objects, and proposes a resistance spot welding quality online detection method with dynamic current and resistance data based on a combined convolutional neural network (CNN), long short-term memory network (LSTM), and an attention mechanism. First, an overall online detection framework using an edge–cloud collaboration was proposed. Second, an online quality detection model was established. In it, the combined CNN and LSTM network were used to extract local detail features and temporal correlation features of the data. The attention mechanism was introduced to improve the interpretability of the model. Moreover, the imbalanced data problem was also solved with a multiclass imbalance algorithm and weighted cross-entropy loss function. Finally, an experimental verification and analysis were conducted. The results show that the quality detection accuracy was 98.5%. The proposed method has good detection performance and real-time detection abilities for the in-site welding processes of automobile bodies. [ABSTRACT FROM AUTHOR]

Digital images cannot be excluded as part of a popular choice of information representation. Covert information can be easily hidden using images. Several schemes are available to hide covert information and are known as steganography schemes. Steganalysis schemes are applied on stego-images to assess the strength of steganography schemes. In this paper, a new steganalysis scheme is presented to detect stego-images. Predefined kernels guide the set of a conventional convolutional layer, and the tight cohesion provides completely guided training. The learning rate of convolutional layers with predefined kernels is higher than the global learning rate. The higher learning rate of the convolutional layers with predefined kernels assures adaptability according to network training, while still maintaining the basic attributes of high-pass kernels. The Leaky ReLU layer is exploited against the ReLU layer to boost the detection performance. Transfer learning is applied to enhance detection performance. The deep network weights are initialized using the weights of the trained network from high-payload stego-images. The strength of the proposed scheme is verified on the HILL, Mi-POD, S-UNIWARD, and WOW content-adaptive steganography schemes. The results are overwhelming and better than the existing steganalysis schemes. [ABSTRACT FROM AUTHOR]

With the emergence of COVID-19 disease in 2019, machine learning (ML) techniques, specifically deep learning networks (DNNs), played a key role in diagnosing the disease in the medical industry due to their superior performance. However, the computational cost of deep learning networks (DNNs) can be quite high, making it necessary to often outsource the training process to third-party providers, such as machine learning as a service (MLaaS). Therefore, careful consideration is required to achieve robustness in DNN-based systems against cyber-security attacks. In this paper, we propose a method called the dropout-bagging (DB-COVIDNet) algorithm, which works as a robust defense mechanism against poisoning backdoor attacks. In this model, the trigger-related features will be removed by the modified dropout algorithm, and then we will use the new voting method in the bagging algorithm to achieve the final results. We considered AC-COVIDNet as the main inducer of the bagging algorithm, which is an attention-guided contrastive convolutional neural network (CNN), and evaluated the performance of the proposed method with the malicious COVIDx dataset. The results demonstrated that DB-COVIDNet has strong robustness and can significantly reduce the effect of the backdoor attack. The proposed DB-COVIDNet nullifies backdoors before the attack has been activated, resulting in a tremendous reduction in the attack success rate from  99.5 %  to  3 %  with high accuracy on the clean data. [ABSTRACT FROM AUTHOR]

The main challenge with rotating machine fault diagnosis is the condition monitoring of machines undergoing nonstationary operations. One possible way of efficiently handling this situation is to use the deep learning (DL) method. However, most DL methods have difficulties when the issue of imbalanced datasets occurs. This paper proposes a novel framework to mitigate this issue by developing an area-metric-based sampling method. In the proposed process, the new sampling scheme can identify which locations of the datasets can potentially have a high degree of surprise. The basic idea of the proposed method is whenever significant deviations from the area metrics are observed to populate more sample points. In addition, to improve the training accuracy of the DL method, the obtained sampled datasets are transformed into a continuous wavelet transform (CWT)-based scalogram representing the time–frequency component. The dilated convolutional neural network (CNN) is also introduced as a classification process with the altered images. The efficacy of the proposed method is demonstrated with fault diagnosis problems for welding robots. The obtained results are also compared with existing methods. [ABSTRACT FROM AUTHOR]

Website fingerprinting attacks attempt to apply deep learning technology to identify websites corresponding to encrypted traffic data. Unfortunately, to the best of our knowledge, once the total number of encrypted traffic data becomes insufficient, the identification accuracy in most existing works will drop dramatically. This phenomenon grows worse because the statistical features of the encrypted traffic data are not always stable but irregularly varying in different time periods. Even a deep learning model requires good performance to capture the statistical features, its accuracy usually diminishes in a short period of time because the changes of the statistical features technically put the training and testing data into two non-identical distributions. In this paper, we first propose a convolutional neural network-based website fingerprinting attack (CWFA) scheme. This scheme integrates packet direction with the timing sequence from the encrypted traffic data to improve the accuracy of analysis as much as possible on few data samples. We then design a new fine-tuning mechanism for the CWFA (FM-CWFA) scheme based on transfer learning. This mechanism enables the proposed FM-CWFA scheme to support the changes in the statistical patterns. The experimental results in closed-world and open-world settings show that the effectiveness of the CWFA scheme is better than previous researches, with the slowest performance degradation when the number of data decreases, and the FM-CWFA scheme can remain effective when the statistical features change. [ABSTRACT FROM AUTHOR]

Currently, the task of maintaining cybersecurity and reliability in various computer systems is relevant. This problem can be solved by detecting anomalies in the log data, which are represented as a stream of textual descriptions of events taking place. For these purposes, reduction to a One-class classification problem is used. Standard One-class classification methods do not achieve good results. Deep learning approaches are more effective. However, they are not robust to outliers and require a lot of computational effort. In this paper, we propose a new robust approach based on a convolutional autoencoder using fuzzy clustering. The proposed approach uses a parallel convolution operation to feature extraction, which makes it more efficient than the currently popular Transformer architecture. In the course of the experiments, the proposed approach showed the best results for both the cybersecurity and the reliability problems compared to existing approaches. It was also shown that the proposed approach is robust to outliers in the training set. [ABSTRACT FROM AUTHOR]

Deep learning has achieved great successes in performing many visual recognition tasks, including object detection. Nevertheless, existing deep networks are computationally expensive and memory intensive, hindering their deployment in resource-constrained environments, such as mobile or embedded devices that are widely used by city travellers. Recently, estimating city-level travel patterns using street imagery has been shown to be a potentially valid way according to a case study with Google Street View (GSV), addressing a critical challenge in transport object detection. This paper presents a compressed deep network using tensor decomposition to detect transport objects in GSV images, which is sustainable and eco-friendly. In particular, a new dataset named Transport Mode Share-Tokyo (TMS-Tokyo) is created to serve the public for transport object detection. This is based on the selection and filtering of 32,555 acquired images that involve 50,827 visible transport objects (including cars, pedestrians, buses, trucks, motors, vans, cyclists and parked bicycles) from the GSV imagery of Tokyo. Then a compressed convolutional neural network (termed SVDet) is proposed for street view object detection via tensor train decomposition on a given baseline detector. The method proposed herein yields a mean average precision (mAP) of 77.6% on the newly introduced dataset, TMS-Tokyo, necessitating just 17.29 M parameters and a computational capacity of 16.52 G FLOPs. As such, it markedly surpasses the performance of existing state-of-the-art methods documented in the literature. [ABSTRACT FROM AUTHOR]

With the advent of smart mobile devices, end users get used to transmitting and storing their individual privacy in them, which, however, has aroused prominent security concerns inevitably. In recent years, numerous researchers have primarily proposed to utilize motion sensors to explore implicit authentication techniques. Nonetheless, for them, there are some significant challenges in real-world scenarios. For example, depending on the expert knowledge, the authentication accuracy is relatively low due to some difficulties in extracting user micro features, and noisy labels in the training phrase. To this end, this paper presents a real-time sensor-based mobile user authentication approach, ST-SVD, a semi-supervised Teacher–Student (TS) tri-training algorithm, and a system with client–server (C-S) architecture. (1) With S-transform and singular value decomposition (ST-SVD), we enhance user micro features by transforming time-series signals into 2D time-frequency images. (2) We employ a Teacher–Student Tri-Training algorithm to reduce label noise within the training sets. (3) To obtain a set of robust parameters for user authentication, we input the well-labeled samples into a CNN (convolutional neural network) model, which validates our proposed system. Experimental results on large-scale datasets show that our approach achieves authentication accuracy of 96.32%, higher than the existing state-of-the-art methods. [ABSTRACT FROM AUTHOR]

The accurate classification of seizure types using electroencephalography (EEG) signals plays a vital role in determining a precise treatment plan and therapy for epilepsy patients. Among the available deep network models, Convolutional Neural Networks (CNNs) are the most widely adopted models for learning and representing EEG signals. However, typical CNNs have high computational complexity, leading to overfitting problems. This paper proposes the design of two effective, lightweight deep network models; the 1D multiscale neural network (1D-MSCNet) model and the Long Short-term Memory (LSTM)-based compact CNN (EEG-LSTMNet) model. The 1D-MSCNet model comprises three modules: a spectral–temporal convolution module, a spatial convolution module, and a classification module. It extracts features from input EEG trials at multiple frequency/time ranges, identifying relationships between the spatial distribution of their channels. The EEG-LSTMNet model includes three convolutional layers, namely temporal, depthwise, and separable layers, a single LSTM layer, and two fully connected classification layers to extract discriminative EEG feature representations. Both models have been applied to the same EEG trials collected from the Temple University Hospital (TUH) database. Results revealed F1-score values of 96.9% and 98.4% for the 1D-MSCNet and EEG-LSTMNet, respectively. Based on the demonstrated outcomes, both models outperform related state-of-the-art methods due to their architectures' adoption of 1D modules and layers that reduce the computational effort needed, solve the overfitting problem, and enhance classification efficiency. Hence, both models could be valuable additions for neurologists to help them decide upon precise treatments and drugs for patients depending on their type of seizure. [ABSTRACT FROM AUTHOR]

This paper concerns the development of a machine learning tool to detect anomalies in the molecular structure of Gallium Arsenide. We employ a combination of a CNN and a PCA reconstruction to create the model, using real images taken with an electron microscope in training and testing. The methodology developed allows for the creation of a defect detection model, without any labeled images of defects being required for training. The model performed well on all tests under the established assumptions, allowing for reliable anomaly detection. To the best of our knowledge, such methods are not currently available in the open literature; thus, this work fills a gap in current capabilities. [ABSTRACT FROM AUTHOR]

The reconfigurable intelligent surface (RIS) is one of the most innovative and revolutionary technologies for increasing the effectiveness of wireless systems. Deep learning (DL) is a promising method that can enhance system efficacy using powerful tools in RIS-based environments. However, the lack of extensive training of the DL model results in the reduced prediction of feature information and performance failure. Hence, to address the issues, in this paper, a combined DL-based optimal decoding model is proposed to improve the transmission error rate and enhance the overall efficiency of the RIS-assisted multiple-input multiple-output communication system. The proposed DL model is comprised of a 1-dimensional convolutional neural network (1-D CNN) and a gated recurrent unit (GRU) module where the 1-D CNN model is employed for the extraction of features from the received signal with further process over the configuration of different layers. Thereafter, the processed data are used by the GRU module for successively retrieving the transmission signal with a minimal error rate and accelerating the convergence rate. It is initially trained offline using created OFDM data sets, after which it is used online to track the channel and extract the transmitted data. The simulation results show that the proposed network performs better than the other technique that was previously used in terms of bit error rate and symbol error rate. The outcomes of the model demonstrate the suitability of the proposed model for use with the next-generation wireless communication system. [ABSTRACT FROM AUTHOR]

When a machine tool is used for a long time, its bearing experiences wear and failure due to heat and vibration, resulting in damage to the machine tool. In order to make the machine tool stable for processing, this paper proposes a smart bearing diagnosis system (SBDS), which uses a gradient-weighted class activation mapping (Grad-CAM)-based convolutional neuro-fuzzy network (GC-CNFN) to detect the bearing status of the machine tool. The developed GC-CNFN is composed of a convolutional layer and neuro-fuzzy network. The convolutional layer can automatically extract vibration signal features, which are then classified using the neuro-fuzzy network. Moreover, Grad-CAM is used to analyze the attention of the diagnosis model. To verify the performance of bearing fault classification, the 1D CNN (ODCNN) and improved 1D LeNet-5 (I1DLeNet) were adopted to compare with the proposed GC-CNFN. Experimental results showed that the proposed GC-CNFN required fewer parameters (20K), had a shorter average calculation time (117.7 s), and had a higher prediction accuracy (99.88%) in bearing fault classification. The proposed SBDS can not only accurately classify bearing faults, but also help users understand the current status of the machine tool. [ABSTRACT FROM AUTHOR]

In recent years, churn rates in industries such as finance have increased, and the cost of acquiring new users is more than five times the cost of retaining existing users. To improve the intelligent prediction accuracy of customer churn rate, artificial intelligence is gradually used. In this paper, the bidirectional long short-term memory convolutional neural network (BiLSTM-CNN) model is integrated with recurrent neural networks (RNNs) and convolutional neural networks (CNNs) in parallel, which well solves the defective problem that RNNs and CNNs run separately, and it also solves the problem that the output results of a long short-term memory network (LSTM) layer in a densely-connected LSTM-CNN (DLCNN) model will ignore some local information when input to the convolutional layer. In order to explore whether the attention bidirectional long short-term memory convolutional neural network (AttnBLSTM-CNN) model can perform better than BiLSTM-CNN, this paper uses bank data to compare the two models. The experimental results show that the accuracy of the AttnBiLSTM-CNN model is improved by 0.2%, the churn rate is improved by 1.3%, the F1 value is improved by 0.0102, and the AUC value is improved by 0.0103 compared with the BLSTM model. Therefore, introducing the attention mechanism into the BiLSTM-CNN model can further improve the performance of the model. The improvement of the accuracy of the user churn prediction model can warn of the possibility of user churn in advance and take effective measures in advance to prevent user churn and improve the core competitiveness of financial institutions. [ABSTRACT FROM AUTHOR]

Image deblurring attracts research attention in the field of image processing and computer vision. Traditional deblurring methods based on statistical prior largely depend on the selected prior type, which limits their restoring ability. Moreover, the constructed deblurring model is difficult to solve, and the operation is comparatively complicated. Meanwhile, deep learning has become a hotspot in various fields in recent years. End-to-end convolutional neural networks (CNNs) can learn the pixel mapping relationships between degraded images and clear images. In addition, they can also obtain the result of effectively eliminating spatial variable blurring. However, conventional CNNs have some disadvantages in generalization ability and details of the restored image. Therefore, this paper presents an iterative dual CNN called IDC for image deblurring, where the task of image deblurring is divided into two sub-networks: deblurring and detail restoration. The deblurring sub-network adopts a U-Net structure to learn the semantical and structural features of the image, and the detail restoration sub-network utilizes a shallow and wide structure without downsampling, where only the image texture features are extracted. Finally, to obtain the deblurred image, this paper presents a multiscale iterative strategy that effectively improves the robustness and precision of the model. The experimental results showed that the proposed method has an excellent effect of deblurring on a real blurred image dataset and is suitable for various real application scenes. [ABSTRACT FROM AUTHOR]

This paper proposes a new linear combination model to predict the closing prices on multivariate financial data sets. The new approach integrates two delays of deep learning methods called the two-delay combination model. The forecasts are derived from three different deep learning models: the multilayer perceptron (MLP), the convolutional neural network (CNN) and the long short-term memory (LSTM) network. Moreover, the weight combination of our proposed model is estimated using the differential evolution (DE) algorithm. The proposed model is built and tested for three high-frequency stock data in financial markets—Microsoft Corporation (MSFT), Johnson & Johnson (JNJ) and Pfizer Inc. (PFE). The individual and combination forecast methods are compared using the root mean square error (RMSE) and the mean absolute percentage error (MAPE). The state-of-the-art combination models used in this paper are the equal weight (EW), the inverse of RMSE (INV-RMSE) and the variance-no-covariance (VAR-NO-CORR) methods. These comparisons demonstrate that our proposed approach using DE weight's optimization has significantly lower forecast errors than the individual model and the state-of-the-art weight combination procedures for all experiments. Consequently, combining two delay deep learning models using differential evolution weights can effectively improve the stock price prediction. [ABSTRACT FROM AUTHOR]

In the increasingly complex domain of Korean voice phishing attacks, advanced detection techniques are paramount. Traditional methods have achieved some degree of success. However, they often fail to detect sophisticated voice phishing attacks, highlighting an urgent need for enhanced approaches to improve detection performance. Addressing this, we have designed and implemented a novel artificial neural network (ANN) architecture that successfully combines data-centric and model-centric AI methodologies for detecting Korean voice phishing attacks. This paper presents our unique hybrid architecture, consisting of a 1-dimensional Convolutional Neural Network (1D CNN), a Bidirectional Long Short-Term Memory (BiLSTM), and Hierarchical Attention Networks (HANs). Our evaluations using the real-world KorCCVi v2 dataset demonstrate that the proposed architecture effectively leverages the strengths of CNN and BiLSTM to extract and learn contextually rich features from word embedding vectors. Additionally, implementing word and sentence attention mechanisms from HANs enhances the model's focus on crucial features, considerably improving detection performance. Achieving an accuracy score of 99.32% and an F1 score of 99.31%, our model surpasses all baseline models we trained, outperforms several existing solutions, and maintains comparable performance to others. The findings of this study underscore the potential of hybrid neural network architectures in improving voice phishing detection in the Korean language and pave the way for future research. This could involve refining and expanding upon this model to tackle increasingly sophisticated voice phishing strategies effectively or utilizing larger datasets. [ABSTRACT FROM AUTHOR]

Success in student learning is the primary aim of the educational system. Artificial intelligence utilizes data and machine learning to achieve excellence in student learning. In this paper, we exploit several machine learning techniques to estimate early student performance. Two main simulations are used for the evaluation. The first simulation used the Traditional Machine Learning Classifiers (TMLCs) applied to the House dataset, and they are Gaussian Naïve Bayes (GNB), Support Vector Machine (SVM), Decision Tree (DT), Multi-Layer Perceptron (MLP), Random Forest (RF), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA). The best results were achieved with the MLP classifier with a division of 80% training and 20% testing, with an accuracy of 88.89%. The fusion of these seven classifiers was also applied and the highest result was equal to the MLP. Moreover, in the second simulation, the Convolutional Neural Network (CNN) was utilized and evaluated on five main datasets, namely, House, Western Ontario University (WOU), Experience Application Programming Interface (XAPI), University of California-Irvine (UCI), and Analytics Vidhya (AV). The UCI dataset was subdivided into three datasets, namely, UCI-Math, UCI-Por, and UCI-Fused. Moreover, the AV dataset has three targets which are Math, Reading, and Writing. The best accuracy results were achieved at 97.5%, 99.55%, 98.57%, 99.28%, 99.40%, 99.67%, 92.93%, 96.99%, and 96.84% for the House, WOU, XAPI, UCI-Math, UCI-Por, UCI-Fused, AV-Math, AV-Reading, and AV-Writing datasets, respectively, under the same protocol of evaluation. The system demonstrates that the proposed CNN-based method surpasses all seven conventional methods and other state-of-the-art-work. [ABSTRACT FROM AUTHOR]

The advancement of 6G networks is driven by the need for customer-centric communication and network control, particularly in applications such as intelligent transport systems. These applications rely on outdoor communication in extremely high-frequency (EHF) bands, including millimeter wave (mmWave) frequencies exceeding 30 GHz. However, EHF signals face challenges such as higher attenuation, diffraction, and reflective losses caused by obstacles in outdoor environments. To overcome these challenges, 6G networks must focus on system designs that enhance propagation characteristics by predicting and mitigating diffraction, reflection, and scattering losses. Strategies such as proper handovers, antenna orientation, and link adaptation techniques based on losses can optimize the propagation environment. Among the network components, aerial networks, including unmanned aerial vehicles (UAVs) and electric vertical take-off and landing aircraft (eVTOL), are particularly susceptible to diffraction losses due to surrounding buildings in urban and suburban areas. Traditional statistical models for estimating the height of tall objects like buildings or trees are insufficient for accurately calculating diffraction losses due to the dynamic nature of user mobility, resulting in increased latency unsuitable for ultra-low latency applications. To address these challenges, this paper proposes a deep learning framework that utilizes easily accessible Google Street View imagery to estimate building heights and predict diffraction losses across various locations. The framework enables real-time decision-making to improve the propagation environment based on users' locations. The proposed approach achieves high accuracy rates, with an accuracy of 39% for relative error below 2%, 83% for relative error below 4%, and 96% for both relative errors below 7% and 10%. Compared to traditional statistical methods, the proposed deep learning approach offers significant advantages in height prediction accuracy, demonstrating its efficacy in supporting the development of 6G networks. The ability to accurately estimate heights and map diffraction losses before network deployment enables proactive optimization and ensures real-time decision-making, enhancing the overall performance of 6G systems. [ABSTRACT FROM AUTHOR]

In recent times, the power sector has become a focal point of extensive scientific interest, driven by a convergence of factors, such as mounting global concerns surrounding climate change, the persistent increase in electricity prices within the wholesale energy market, and the surge in investments catalyzed by technological advancements across diverse sectors. These evolving challenges have necessitated the emergence of new imperatives aimed at effectively managing energy resources, ensuring grid stability, bolstering reliability, and making informed decisions. One area that has garnered particular attention is the accurate prediction of end-user electricity load, which has emerged as a critical facet in the pursuit of efficient energy management. To tackle this challenge, machine and deep learning models have emerged as popular and promising approaches, owing to their having remarkable effectiveness in handling complex time series data. In this paper, the development of an algorithmic model that leverages an automated process to provide highly accurate predictions of electricity load, specifically tailored for the island of Thira in Greece, is introduced. Through the implementation of an automated application, an array of deep learning forecasting models were meticulously crafted, encompassing the Multilayer Perceptron, Long Short-Term Memory (LSTM), One Dimensional Convolutional Neural Network (CNN-1D), hybrid CNN–LSTM, Temporal Convolutional Network (TCN), and an innovative hybrid model called the Convolutional LSTM Encoder–Decoder. Through evaluation of prediction accuracy, satisfactory performance across all the models considered was observed, with the proposed hybrid model showcasing the highest level of accuracy. These findings underscore the profound significance of employing deep learning techniques for precise forecasting of electricity demand, thereby offering valuable insights with which to tackle the multifaceted challenges encountered within the power sector. By adopting advanced forecasting methodologies, the electricity sector moves towards greater efficiency, resilience and sustainability. [ABSTRACT FROM AUTHOR]

Artificial intelligence plays a significant role in traffic-accident detection. Traffic accidents involve a cascade of inadvertent events, making traditional detection approaches challenging. For instance, Convolutional Neural Network (CNN)-based approaches cannot analyze temporal relationships among objects, and Recurrent Neural Network (RNN)-based approaches suffer from low processing speeds and cannot detect traffic accidents simultaneously across multiple frames. Furthermore, these networks dismiss background interference in input video frames. This paper proposes a framework that begins by subtracting the background based on You Only Look Once (YOLOv5), which adaptively reduces background interference when detecting objects. Subsequently, the CNN encoder and Transformer decoder are combined into an end-to-end model to extract the spatial and temporal features between different time points, allowing for a parallel analysis between input video frames. The proposed framework was evaluated on the Car Crash Dataset through a series of comparison and ablation experiments. Our framework was benchmarked against three accident-detection models to evaluate its effectiveness, and the proposed framework demonstrated a superior accuracy of approximately 96%. The results of the ablation experiments indicate that when background subtraction was not incorporated into the proposed framework, the values of all evaluation indicators decreased by approximately 3%. [ABSTRACT FROM AUTHOR]

Deep learning models with convolutional operators have received widespread attention for their good image denoising performance. However, since the convolutional operation prefers to extract local features, the extracted features may lose some global information, such as texture, structure, and color characteristics, when the object in the image is large. To address this issue, this paper proposes an N-shaped convolutional neural network with the ability to extract multi-scale features to capture more useful information and alleviate the problem of global information loss. The proposed network has two main parts: a multi-scale input layer and a multi-scale feature extraction layer. The former uses a two-dimensional Haar wavelet to create an image pyramid, which contains the corrupted image's high- and low-frequency components at different scales. The latter uses a U-shaped convolutional network to extract features at different scales from this image pyramid. The method sets the mean-squared error as the loss function and uses the residual learning strategy to learn the image noise directly. Compared with some existing image denoising methods, the proposed method shows good performance in gray and color image denoising, especially in textures and contours. [ABSTRACT FROM AUTHOR]

The purpose of this study is to investigate the role that a deep learning approach could play in computational mechanics. In this paper, a convolutional neural network technique based on modified loss function is proposed as a surrogate of the finite element method (FEM). Several surrogate-based physics-informed neural networks (PINNs) are developed to solve representative boundary value problems based on elliptic partial differential equations (PDEs). According to the authors' knowledge, the proposed method has been applied for the first time to solve boundary value problems with elliptic partial differential equations as the governing equations. The results of the proposed surrogate-based approach are in good agreement with those of the conventional FEM. It is found that modification of the loss function could improve the prediction accuracy of the neural network. It is demonstrated that to some extent, the deep learning approach could replace the conventional numerical method as a significant surrogate model. [ABSTRACT FROM AUTHOR]

Accurately identifying the seizure type is vital in the treatment plan and drug prescription for epileptic patients. The most commonly adopted test for identifying epileptic seizures is electroencephalography (EEG). EEG signals include important information about the brain's electrical activities and are widely used for epilepsy analysis. Among various deep network architectures, convolutional neural networks (CNNs) have been widely used for EEG signal representation learning for epilepsy analysis. However, most of the existing CNN-based methods suffer from the overfitting problem due to a small number of EEG trials and the huge number of learnable parameters. This paper introduces the design of an efficient, lightweight, and expressive deep network model based on ResNet theory and long short-term memory (LSTM) for classifying seizure types from EEG trials. A 1D ResNet module is adopted to train a deeper network without encountering vanishing gradient problems and to avoid the overfitting problem of CNN models. The LSTM module encodes and learns long-term dependencies over time. The synthetic minority oversampling technique (SMOTE) is applied to balance the data by increasing the trials of minority classes. The proposed method was evaluated using the public domain benchmark TUH database. Experimental results revealed the superior performance of the proposed model over other state-of-the-art models with an F1-score of 97.4%. The proposed deep learning model will help neurologists precisely interpret and classify epileptic seizure types and enhance the patient's life. [ABSTRACT FROM AUTHOR]

Accurate stock price prediction has an important role in stock investment. Because stock price data are characterized by high frequency, nonlinearity, and long memory, predicting stock prices precisely is challenging. Various forecasting methods have been proposed, from classical time series methods to machine-learning-based methods, such as random forest (RF), recurrent neural network (RNN), convolutional neural network (CNN), Long Short-Term Memory (LSTM) neural networks and their variants, etc. Each method can reach a certain level of accuracy but also has its limitations. In this paper, a CNN-BiLSTM-Attention-based model is proposed to boost the accuracy of predicting stock prices and indices. First, the temporal features of sequence data are extracted using a convolutional neural network (CNN) and bi-directional long and short-term memory (BiLSTM) network. Then, an attention mechanism is introduced to fit weight assignments to the information features automatically; and finally, the final prediction results are output through the dense layer. The proposed method was first used to predict the price of the Chinese stock index—the CSI300 index and was found to be more accurate than any of the other three methods—LSTM, CNN-LSTM, CNN-LSTM-Attention. In order to investigate whether the proposed model is robustly effective in predicting stock indices, three other stock indices in China and eight international stock indices were selected to test, and the robust effectiveness of the CNN-BiLSTM-Attention model in predicting stock prices was confirmed. Comparing this method with the LSTM, CNN-LSTM, and CNN-LSTM-Attention models, it is found that the accuracy of stock price prediction is highest using the CNN-BiLSTM-Attention model in almost all cases. [ABSTRACT FROM AUTHOR]

With the development of artificial intelligence, techniques such as machine learning, object detection, and trajectory tracking have been applied to various traffic fields to detect accidents and analyze their causes. However, detecting traffic accidents using closed-circuit television (CCTV) as an emerging subject in machine learning remains challenging because of complex traffic environments and limited vision. Traditional research has limitations in deducing the trajectories of accident-related objects and extracting the spatiotemporal relationships among objects. This paper proposes a traffic accident detection method that helps to determine whether each frame shows accidents by generating and considering object trajectories using influence maps and a convolutional neural network (CNN). The influence maps with spatiotemporal relationships were enhanced to improve the detection of traffic accidents. A CNN is utilized to extract latent representations from the influence maps produced by object trajectories. Car Accident Detection and Prediction (CADP) was utilized in the experiments to train our model, which achieved a traffic accident detection accuracy of approximately 95%. Thus, the proposed method attained remarkable results in terms of performance improvement compared to methods that only rely on CNN-based detection. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel creation method of feature graphics by deep learning algorithms based on a channel attention module consisting of a separable deep convolutional neural network and an SENet network. The main innovation of this method is that the image feature of sample images is extracted by convolution operation and the key point matrix is obtained by channel weighting calculation to create feature graphics within the channel attention module. The main problem of existing image generation methods is that the complex network training and calculation process affects the accuracy and efficiency of image generation. It greatly reduced the complexity of image generation and improved the efficiency when we trained the image generation network with the feature graphic maps. To verify the superiority of this method, we conducted a comparative experiment with the existing method. Additionally, we explored the influence on the accuracy and efficiency of image generation of the channel number of the weighting matrix based on the test experiment. The experimental results demonstrate that this method highlights the image features of geometric lines, simplifies the complexity of image generation and improves the efficiency. Based on this method, images with more prominent line features are generated from the description text and dynamic graphics are created for the display of the images generated, which can be applied in the construction of smart museums. [ABSTRACT FROM AUTHOR]

The collar is an important part of a garment that reflects its style. The collar classification task is to recognize the collar type in the apparel image. In this paper, we design a novel convolutional module called MFA (multi-scale features attention) to address the problems of high noise, small recognition target and unsatisfactory classification effect in collar feature recognition, which first extracts multi-scale features from the input feature map and then encodes them into an attention weight vector to enhance the representation of important parts, thus improving the ability of the convolutional block to combat noise and extract small target object features. It also reduces the computational overhead of the MFA module by using the depth-separable convolution method. Experiments on the collar dataset Collar6 and the apparel dataset DeepFashion6 (a subset of the DeepFashion database) show that MFANet is able to perform at a relatively small number of collars. MFANet can achieve better classification performance than most current mainstream convolutional neural networks for complex collar images with less computational overhead. Experiments on the standard dataset CIFAR-10 show that MFANet also outperforms current mainstream image classification algorithms. [ABSTRACT FROM AUTHOR]

Traffic closed-circuit television (CCTV) devices can be used to detect and track objects on roads by designing and applying artificial intelligence and deep learning models. However, extracting useful information from the detected objects and determining the occurrence of traffic accidents are usually difficult. This paper proposes a CCTV frame-based hybrid traffic accident classification model that enables the identification of whether a frame includes accidents by generating object trajectories. The proposed model utilizes a Vision Transformer (ViT) and a Convolutional Neural Network (CNN) to extract latent representations from each frame and corresponding trajectories. The fusion of frame and trajectory features was performed to improve the traffic accident classification ability of the proposed hybrid method. In the experiments, the Car Accident Detection and Prediction (CADP) dataset was used to train the hybrid model, and the accuracy of the model was approximately 97%. The experimental results indicate that the proposed hybrid method demonstrates an improved classification performance compared to traditional models. [ABSTRACT FROM AUTHOR]

Image super-resolution (SR), as one of the classic image processing issues, has attracted increasing attention from researchers. As a highly ill-conditioned, non-convex optimization issue, it is difficult for image SR to restore a high-resolution (HR) image from a given low-resolution (LR) instance. Recent researchers have tended to regard image SR as a regression task and to design an end-to-end convolutional neural network (CNN) to predict the pixels directly, which lacks inherent theoretical analysis and limits the effectiveness of the restoration. In this paper, we analyze image SR from an optimization perspective and develop a deep successive convex approximation network (SCANet) for generating HR images. Specifically, we divide non-convex optimization into several convex LASSO sub-problems and use CNN to adaptively learn the parameters. To boost network representation, we use residual feature aggregation (RFA) blocks and devise a spatial and channel attention (SACA) mechanism to improve the restoration capacity. The experimental results show that the proposed SCANet can restore HR images more effectively than other works. Specifically, SCANet achieves higher PSNR/SSIM results and generates more satisfying textures. [ABSTRACT FROM AUTHOR]

An automatic temporal video segmentation framework is introduced in this article. The proposed cut detection technique performs a high-level feature extraction on the video frames, by applying a multi-scale image analysis approach combining nonlinear partial differential equations (PDE) to convolutional neural networks (CNN). A nonlinear second-order hyperbolic PDE model is proposed and its well-posedness is then investigated rigorously here. Its weak and unique solution is determined numerically applying a finite difference method-based numerical approximation algorithm that quickly converges to it. A scale-space representation is then created using that iterative discretization scheme. A CNN-based feature extraction is performed at each scale and the feature vectors obtained at multiple scales are concatenated into a final frame descriptor. The feature vector distance values between any two successive frames are then determined and the video transitions are identified next, by applying an automatic clustering scheme on these values. The proposed PDE model, its mathematical investigation and discretization, and the multi-scale analysis based on it represent the major contributions of this work. Some temporal segmentation experiments and method comparisons that illustrate the effectiveness of the proposed framework are finally described in this research paper. [ABSTRACT FROM AUTHOR]

Writer recognition based on a small amount of handwritten text is one of the most challenging deep learning problems because of the implicit characteristics of handwriting styles. In a deep convolutional neural network, writer recognition based on supervised learning has shown great success. These supervised methods typically require a lot of annotated data. However, collecting annotated data is expensive. Although unsupervised writer recognition methods may address data annotation issues significantly, they often fail to capture sufficient feature relationships and usually perform less efficiently than supervised learning methods. Self-supervised learning may solve the unlabeled dataset issue and train the unsupervised datasets in a supervised manner. This paper introduces Self-Writer, a self-supervised writer recognition approach dealing with unlabeled data. The proposed scheme generates clusterable embeddings from a small fixed-length image frame such as a text block. The training strategy presumes that a small image frame of handwritten text should include the writer's handwriting characteristics. We construct pairwise constraints and nongenerative augmentation to train Siamese architecture to generate embeddings depending on such an assumption. Self-Writer is evaluated on the two most widely used datasets, IAM and CVL, on pairwise and triplet architecture. We find Self-Writer to be convincing in achieving satisfactory performance using pairwise architectures. [ABSTRACT FROM AUTHOR]

Due to the COVID-19 pandemic, the necessity for a contactless biometric system able to recognize masked faces drew attention to the periocular region as a valuable biometric trait. However, periocular recognition remains challenging for deployments in the wild or in unconstrained environments where images are captured under non-ideal conditions with large variations in illumination, occlusion, pose, and resolution. These variations increase within-class variability and between-class similarity, which degrades the discriminative power of the features extracted from the periocular trait. Despite the remarkable success of convolutional neural network (CNN) training, CNN requires a huge volume of data, which is not available for periocular recognition. In addition, the focus is on reducing the loss between the actual class and the predicted class but not on learning the discriminative features. To address these problems, in this paper we used a pre-trained CNN model as a backbone and introduced an effective deep CNN periocular recognition model, called linear discriminant analysis CNN (LDA-CNN), where an LDA layer was incorporated after the last convolution layer of the backbone model. The LDA layer enforced the model to learn features so that the within-class variation was small, and the between-class separation was large. Finally, a new fully connected (FC) layer with softmax activation was added after the LDA layer, and it was fine-tuned in an end-to-end manner. Our proposed model was extensively evaluated using the following four benchmark unconstrained periocular datasets: UFPR, UBIRIS.v2, VISOB, and UBIPr. The experimental results indicated that LDA-CNN outperformed the state-of-the-art methods for periocular recognition in unconstrained environments. To interpret the performance, we visualized the discriminative power of the features extracted from different layers of the LDA-CNN model using the t-distributed Stochastic Neighboring Embedding (t-SNE) visualization technique. Moreover, we conducted cross-condition experiments (cross-light, cross-sensor, cross-eye, cross-pose, and cross-database) that proved the ability of the proposed model to generalize well to different unconstrained conditions. [ABSTRACT FROM AUTHOR]

The latest achievements in the field of reinforcement learning have encouraged the development of vision-based learning methods that compete with human-provided results obtained on various games and training environments. Convolutional neural networks together with Q-learning-based approaches have managed to solve and outperform human players in environments such as Atari 2600, Doom or StarCraft II, but the niche of 3D realistic games with a high degree of freedom of movement and rich graphics remains unexplored, despite having the highest resemblance to real-world situations. In this paper, we propose a novel testbed to push the limits of deep learning methods, namely an OpenAI Gym-like environment based on Dark Souls III, a notoriously difficult role-playing game, where even human players have reportedly struggled. We explore two types of architectures, Deep Q-Network and Deep Recurrent Q-Network, providing the results of a first incursion into this new problem class. The source code for the training environment and baselines is made available. [ABSTRACT FROM AUTHOR]

Tongue color classification serves as important assistance for traditional Chinese medicine (TCM) doctors to make a precise diagnosis. This paper proposes a novel two-step framework based on deep learning to improve the performance of tongue color classification. First, a semantic-based CNN called SegTongue is applied to segment the tongues from the background. Based on DeepLabv3+, multiple atrous spatial pyramid pooling (ASPP) modules are added, and the number of iterations of fusions of low-level and high-level information is increased. After segmentation, various classical feature extraction networks are trained using softmax and center loss. The experiment results are evaluated using different measures, including overall accuracy, Kappa coefficient, individual sensitivity, etc. The results demonstrate that the proposed framework with SVM achieves up to 97.60% accuracy in the tongue image datasets. [ABSTRACT FROM AUTHOR]

Semantic relationships between words provide relevant information about the whole idea in the texts. Existing embedding representation models characterize each word as a vector of numbers with a fixed length. These models have been used in tasks involving text classification, such as recommendation and question–answer systems. However, the embedded information provided by semantic relationships has been neglected. Therefore, this paper proposes an approach that involves semantic relationships in embedding models for text classification, which is evaluated. Three embedding models based on semantic relations extracted from Wikipedia are presented and compared with existing word-based models. Our approach considers the following relationships: synonymy, hyponymy, and hyperonymy. They were considered since previous experiments have shown that they provide semantic knowledge. The relationships are extracted from Wikipedia using lexical-syntactic patterns identified in the literature. The extracted relationships are embedded as a vector: synonymy, hyponymy–hyperonymy, and a combination of all relationships. A Convolutional Neural Network using semantic relationship embeddings was trained for text classification. An evaluation was carried out for the proposed relationship embedding configurations and existing word-based models to compare them based on two corpora. The results were obtained with the metrics of precision, accuracy, recall, and F 1 -measure. The best results for the 20-Newsgroup corpus were obtained with the hyponymy–hyperonymy embeddings, achieving an accuracy of 0.79. For the Reuters corpus, F 1 -measure and recall of 0.87 were obtained using synonymy–hyponymy–hyperonymy. [ABSTRACT FROM AUTHOR]