Search

In this paper, a novel Extended Logistic Chaotic Map (ELCM) with two control parameters is proposed. Overcoming the major drawback of the standard Logistic and other existing chaotic maps, the ELCM not only has infinite chaotic range as well as good ergodicity, but also has a simple structure just like the Logistic map, which greatly facilitates its practical implementation and becomes very suitable for today's real-time applications. Moreover, a new color image encryption scheme based on chaos concept, DNA (Deoxyribonucleic Acid) encoding and Convolutional Neural Network (CNN) is also presented. To perform a chaotic scrambling, the ELCM is however used at different stages of the encryption process. In addition, a pre-trained AlexNet CNN is used to generate a public key. After XORing with the secret key, the latter is used, on the one hand, to generate the initials values and the control parameters of the ELCM. On the other hand, it will be split into two keys in order to generate a random grayscale image, which will then be XORed with the three components (i.e., R, G and B) of the color plaintext image. Afterward, a permutation operation, DNA encoding, diffusion operation as well as bit reversion operation are then applied to the R, G and B components. The performance evaluation demonstrates that the ELCM not only has an infinite chaotic parameter range, but also exhibits a high chaotic complexity. Besides, experimental results as well as security analysis confirm that with NPCR of 99.6287%, the designed color image encryption scheme achieves an average entropy of 7.9975 and a near zero correlation (-0.0027). Furthermore, the proposed encryption scheme is in fact of higher security level in comparison to other schemes recently presented in the literature, thus making it highly suitable for today's real-time applications. [ABSTRACT FROM AUTHOR]

Brain-computer interface (BCI) provides a platform for humans to communicate using Electroencephalogram (EEG) signals by converting them into commands that can be used by the output device to perform the desired tasks. This paper focuses on the identification of vowels from EEG signals. First, a dataset of EEG signals has been created for the identification of vowels by collecting data using a 14-channel EEG device Emotiv- epoc+ from 16 subjects. Then, a deep learning-based model is proposed using a multi-headed Convolutional Neural Network for feature extraction and classification of imagined speech of vowels. Butterworth lowpass and bandpass filter of order five are implemented for denoising and sub-banding of the EEG signals which are further pre-processed using Hilbert Huang Transform. The model has achieved an average accuracy of 97.67% with a five-fold cross-validation technique using all six sub-bands of the EEG signals. The model has achieved an average precision and recall of 95.54% and 95.11% respectively. The proposed model is statistically tested using the Mann-Whitney U test and paired t-test with a p-value less than 0.05. [ABSTRACT FROM AUTHOR]

Recently, deep learning methods have achieved excellent performance in single image super-resolution (SISR), but most existing methods suffer from heavy computational costs and memory storage. To address this problem, some lightweight SR methods have been proposed, among which convolutional neural network (CNN) with attention mechanisms has received increasing attention. However, the existing CNN-based lightweight network and attention mechanism still have redundancy in feature extraction and enhancement. This paper presents Efficient Self-calibrated and Hierarchical Refinement Network (ESHR) for Lightweight Super-Resolution that includes three effective designs. First, we propose a more effective feature extraction basic module called Self-calibrated Residual Block (SRB), which can achieve superior performance when replacing other common convolution modules such as traditional convolution, BSConv and DSConv. Second, we propose a more lightweight spatial attention mechanism called Squeeze-Excitation Spatial Attention (SESA), which can interact and enhance features more effectively in spatial dimensions to improve the representation ability of the model. Third, we design an Efficient Hierarchical Refinement Block (EHRB) to integrate and utilize multi-level features, which can achieve better performance than single-level refinement without increasing the number of parameters. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art CNN-based methods quantitatively and qualitatively. [ABSTRACT FROM AUTHOR]

This paper proposes low-biased probabilistic multipliers and applies the proposed designs to the inference stage of convolutional neural networks (CNNs). Highly inaccurate compressors in the probabilistic multiplier make the error distribution unbalanced and produce significant relative errors. We describe design rules for applying different compressors to the probabilistic multiplier. Besides, the proposed design rules enhance the error characteristics by determining the input values of each compressor, suppressing the significant relative error. Notably, motivated by the design rules, we propose a novel inaccurate 4:2 compressor suitable for our new design. The design rules and proposed 4:2 compressor are adopted in 8-bit probabilistic multipliers. The error analysis shows that the new designs can enhance error characteristics without increasing hardware costs. Furthermore, the proposed approximate multipliers are applied to the convolutional layers of CNNs. We describe the fine-tuning scheme to reduce the retraining time of approximation-aware training. The experiments show that proposed designs can produce acceptable classification results compared with those using the floating-point and 8-bit exact multiplications on CIFAR and ImageNet datasets. [ABSTRACT FROM AUTHOR]

Currently, weapon detection through video surveillance has been extensively studied using deep learning techniques by researchers. However, limited attention has been given to the detection of weapons in night time or dark scenarios. This paper aims to address this gap by proposing a novel approach for weapon detection specifically modified for low-light conditions. The authors demonstrate accurate and robust detection of weapons in challenging nighttime environments by modifying the deep learning model YOLOv7. The YOLOv7-DarkVision model has been developed by combining a brightening algorithm with the advanced image processing techniques and architecture of YOLOv7. A dataset of 15,367 images were collected for training of the model, along with five dark videos from various sources for performance evaluation. The derived detection model, which has a precision score of 95.50% and an F1-Score of 93.41%, performs absolutely well as a weapon detector. [ABSTRACT FROM AUTHOR]

Electroencephalogram (EEG) signals contains a major role in examining the behavior of brain activity. Moreover, these signals are contaminated with artifacts, which may affect the EEG signal analysis. The EEG performs as the low spatial resolution in the unimodal analysis. Owing to the highly time-consumption procedure, the removal of artifact techniques cannot be applied in practical applications. In existing works, automatic artifact removal techniques are ineffective for eliminating the artifacts in the EEG signals. Thus, the core of this concept is to design a low complexity aware, intelligent artifacts removal model using deep learning. This paper attempts to introduce a 1-Dimensional-Convolutional Neural Network with a Weighted Deep Feature (1D-CNN-WDF), which acts as a filter to remove the unnecessary noise in EEG signals. The WDF extraction in 1D-CNN-WDF is developed by combining the Coyote Optimization Algorithm (COA), and Bird Swarm Algorithm (BSA) named Hybrid Bird-COA (HB-COA). The same hybrid algorithm is used for the tuning of model parameters in 1D-CNN-WDF. The analysis results confirm that the recommended method is more enriched than recently introduced methods when considering the improved Signal to Noise Ratio (SNR), less Mean Squared Error (MSE), and reduced computational complexity, which shows that the suggested method is attained the superior reconstruction quality. [ABSTRACT FROM AUTHOR]

Near-field multiple-input multiple-output (MIMO) radar imaging systems have recently gained significant attention. These systems generally reconstruct the three-dimensional (3D) complex-valued reflectivity distribution of the scene using sparse measurements. Consequently, imaging quality highly relies on the image reconstruction approach. Existing analytical reconstruction approaches suffer from either high computational cost or low image quality. In this paper, we develop novel non-iterative deep learning-based reconstruction methods for real-time near-field MIMO imaging. The goal is to achieve high image quality with low computational cost at compressive settings. The developed approaches have two stages. In the first approach, physics-based initial stage performs adjoint operation to back-project the measurements to the image-space, and deep neural network (DNN)-based second stage converts the 3D backprojected measurements to a magnitude-only reflectivity image. Since scene reflectivities often have random phase, DNN processes directly the magnitude of the adjoint result. As DNN, 3D U-Net is used to jointly exploit range and cross-range correlations. To comparatively evaluate the significance of exploiting physics in a learning-based approach, two additional approaches that replace the physics-based first stage with fully connected layers are also developed as purely learning-based methods. The performance is also analyzed by changing the DNN architecture for the second stage to include complex-valued processing (instead of magnitude-only processing), 2D convolution kernels (instead of 3D), and ResNet architecture (instead of U-Net). Moreover, we develop a synthesizer to generate large-scale dataset for training the neural networks with 3D extended targets. We illustrate the performance through experimental data and extensive simulations. The results show the effectiveness of the developed physics-based learned reconstruction approach compared to commonly used approaches in terms of both runtime and image quality at highly compressive settings. Our source codes and dataset are made available at https://github.com/METU-SPACE-Lab/Efficient-Learned-3D-Near-Field-MIMO-Imaging upon publication to advance research in this field. • 3D image reconstruction methods are developed for near-field MIMO radar imaging using deep learning-based direct inversion and 3D CNNs. • A synthesizer to generate 3D scenes with extended targets is developed to obtain large data for training the neural networks. • Resolution analysis is performed for compressive imaging settings with sparse arrays. • Performance is comprehensively evaluated with simulated and experimental data, using both magnitude-only and complex-valued processing. • The results show the effectiveness of the developed physics-based learned reconstruction method in terms of image quality and runtime. [ABSTRACT FROM AUTHOR]

The traffic driving scenario is dynamically varying and experienced drivers can allocate visual attention effectively in brief time intervals, focusing on prominent targets and areas in advance, thus ensuring safe driving. Therefore, it is significant to research the behavior of driver's visual attention allocation in the development of driver assistance systems and autonomous vehicles. Recently, inspired by top-down and bottom-up attention mechanisms, most visual attention models based on neural networks have been proposed. However, these models tend to yield highly distributed predictions for the specific driving scenario, and have relatively low prediction accuracy, making them unfeasible for being used in practice. Hence, to overcome these challenges, we propose a driver's visual attention model (DVAM) built on the encoder-decoder architecture with the utilization of the convolutional neural network (CNN) and the recurrent neural networks (RNN). Specifically, the model leverages the temporal and spatial dimension information to better mimic realistic dynamic driving, and extract abundant feature information, thus ensuring that the final predictions are authentic and effective. Then, extensive experiments demonstrate that our suggested DVAM achieves better robustness and precision in predicting driver attention to areas or targets compared to state-of-the-art methods on the DR(eye)VE dataset. Finally, our model is deployed on the TDV dataset to validate its generalization capabilities, which prove the potential for field adaptation. • Inspired by the attention mechanism, this paper proposes a hybrid model for driver visual attention prediction. • The model better mimics the authentic driving condition from static and dynamic perspectives. • The model has been training, experimental validation, and generalization exploratory analysis. [ABSTRACT FROM AUTHOR]

This paper proposes a deep learning-based method for image restoration given an inaccurate knowledge of the degradation. We first show how the impulse response of a Wiener filter can approximate the Moore-Penrose pseudo-inverse of the blur convolution operator. The deconvolution problem is then cast as the learning of a residual in the null space of the blur kernel, which, when added to the Wiener restoration, will satisfy the image formation model. This approach is expected to make the network capable of dealing with different blurs since only residuals associated with the Wiener filter have to be learned. Artifacts caused by inaccuracies in the blur estimation and other image formation model inconsistencies are removed by a Dynamic Filter Network. The extensive experiments carried out on several synthetic and real image datasets assert the proposed method's performance and robustness and demonstrate the advantage of the proposed method over existing ones. [ABSTRACT FROM AUTHOR]

Target detection algorithms are an important technology in the field of computer vision and usually need to be implemented on devices such as CPU or GPU with high computational power. However, traditional CPU are unable to support real-time processing and computations; the power consumption of GPU is relatively high and the application scenarios are limited. FPGA, with their high throughput rate, high power efficiency and reconfigurability, are the best platform for research on hardware acceleration of target detection algorithms. Therefore, how to implement efficient target detection algorithms on embedded devices is an important challenge. A method for YOLOv4-Tiny object detection algorithm accelerator based on FPGA is proposed and demonstrated in this paper. In this design, an optimization strategy is proposed, including the design of IP cores for convolutional computation in HLS, the fusion of batch normalization layers with convolutional layers, dynamic fixed-point 16-bit quantization, loop unrolling double-buffered storage and channel parallel acceleration, which can improve the computational performance and efficiency of the target detection algorithm. The YOLOv4-Tiny target detection algorithm was deployed on an FPGA development board, and experimental results showed that it achieved a computational performance of 18.32 GOP/s and energy efficiency of 6.66 GOP-W/s on the FPGA, an increase of 1.81 times in computational performance and 1.86 times in energy efficiency compared to the same platform. Compared to unquantized CPU platforms, the computational performance is 0.51 times better, the energy efficiency is 34 times better, and the latency of accelerating a single image is reduced to 383 ms, which allows for better detection performance in the field of target detection and tracking. [ABSTRACT FROM AUTHOR]

Digital images are playing a progressively important role in almost all the fields such as computer science, medicine, communications, transmission, security, surveillance, and many more. Digital images are susceptible to a number of distortions due to faulty imaging instruments, transmission channels, atmospheric and environmental conditions, etc. resulting in degraded images. Degradation can be of different types such as noise, backscattering, low saturation, low contrast, tilt, spectral absorption, blurring, etc. The degradation reduces digital images' effectiveness and therefore needs to be restored. In this paper, we present an extensive review of image restoration tasks. It addresses problems like image deblurring, denoising, dehazing and super-resolution. Image restoration is fundamentally an image processing problem, but deep learning techniques, based mainly on convolutional neural networks have received a lot of attention in almost all areas of computer science. Along with deep learning, other machine learning methods have also been tried for restoring digital images. In this review, we have therefore categorized digital image restoration techniques as either image processing-based, machine learning-based or deep learning-based. For each category, a variety of approaches presented in recent years have been reviewed. This review also includes a summary of the data sets used for image restoration along with a baseline reference that can be used by future researchers to compare and improve their results. We also suggest some interesting research directions for future work in this area. [ABSTRACT FROM AUTHOR]

Encoding-decoding convolutional neural network structures have shown powerful performance in contour detection tasks, however, the design of models in the past seems to have little reference to physiological mechanisms, especially biological visual mechanisms. And physiological studies have shown that the visual system is efficient and accurate in the extraction of contour features. Meanwhile, when designing network models, researchers pay more attention to the impact of decoding networks on the performance of the models and pay little attention to the impact of encoding networks on the performance of the models. Therefore, in this paper, inspired by the effective mechanism of the biological vision system for contour detection, we combine it with convolutional neural net and self-attention mechanism to design a multi-level interaction contour detection model. We call it MI-Net. Among them, the multi-level interaction module is designed to simulate the transmission mechanism of visual information across the visual cortex and the feedback mechanism of visual information to realize the multi-level interaction of different levels of the encoding network, and then optimize each visual level to achieve the improvement of the performance of contour detection. The feature integration module is designed by simulating the feature integration function of the inferior temporal (IT) cortex. Multiple feature integration modules are combined into a decoding network to increase the feature integration capability of the decoding network. Experiments on publicly available datasets show that our proposed contour detection network has good performance. [ABSTRACT FROM AUTHOR]

Existing steganographic methods based on adversarial images can only design adversarial images for a single steganalyzer and cannot resist detection from the latest steganalyzers using convolutional neural networks such as SRNet and Zhu-Net. To address this issue, this paper proposes a novel method for enhancing the security of image steganography using multiple adversarial networks and channel attention modules. The proposed method employs generative adversarial networks based on the U-Net structure to generate high-quality adversarial images and uses the self-learning properties of the adversarial networks to iteratively optimize the parameters of multiple adversarial steganographic networks. This process generates high-quality adversarial images capable of misleading multiple steganalyzers. Additionally, the proposed scheme adaptively adjusts the distribution of adversarial noise in the original image using multiple lightweight channel attention modules in the generator, thus enhancing the anti-steganalysis ability of adversarial images. Furthermore, the proposed method utilizes multiple discrimination losses and MSE loss, dynamically combined to improve the quality of adversarial images and facilitate the network's rapid and stable convergence. Extensive experimental results demonstrate that the proposed algorithm can generate adversarial images with an average PSNR of 40.3 dB, and the success rate of misleading the advanced steganalyzers is over 93%. The security and generalization of the algorithm we propose exceed those of the compared steganographic methods. [ABSTRACT FROM AUTHOR]

The secondary user (SU) only detects one feature of the primary user (PU) signal in conventional spectrum sensing. However, single feature detection does not fully explore the PU information. Meanwhile, non-Gaussian interference will also affect the spectrum sensing performance. In this paper, we propose a novel spectrum sensing method using multimodal fusion and convolutional neural networks (CNN). First, the received signal at multiple antennas is preprocessed, where the generalized covariance matrix and the generalized Wigner-Ville distribution are used to characterize two different modes of the received signals as the input of CNN. Then, a CNN model using multimodal fusion is constructed. In this model, the attention mechanism is first used to extract features of different modalities, and the multimodal fusion technique fuses the extracted features to obtain the feature vector. Finally, the detection statistics and detection thresholds for spectrum sensing are constructed and compared using the output feature vectors of CNN. Simulation results show that the performance of multi-mode fusion spectrum sensing is better than the single-mode counterpart. This method can effectively realize spectrum sensing with both Gaussian noise and non-Gaussian interference, and it has a high detection probability at low SNR. [ABSTRACT FROM AUTHOR]

As a new type of surveillance technology, Automatic Dependent Surveillance-Broadcast (ADS-B) has been regarded as the cornerstone of the next generation air surveillance. However, a potential problem is that multiple ADS-B signals may overlap when they come from different aircraft simultaneously. This paper explores the possibility of using deep learning to separate pulsed ADS-B overlapping signals. Most existing deep learning algorithms separate the overlapping signals by utilizing the difference in time-frequency domains. But all ADS-B signals are transmitted at the same frequency, those deep learning methods based on time-frequency analysis are ineffective. This paper separates pulsed ADS-B overlapping signals directly in the time domain. Firstly, an encoder is used to extract the mixed features of ADS-B overlapping signals in the time domain. Then, Temporal Convolutional Network (TCN) is used to extract the non-overlapping ADS-B signal feature from the output of encoder. Finally, the separated time-domain signals are reconstructed by a decoder. Simulation results show that the separation accuracy of the algorithm based on deep learning is higher than that of the traditional algorithm. • The effectiveness of deep learning technology to separate pulsed ADS-B overlapping signals is verified. • The training data set of ADS-B overlapping signal is constructed from the parameters in the actual situation. • Simulation results show that the deep learning algorithm performs better than the traditional algorithm. [ABSTRACT FROM AUTHOR]

This paper introduces an approach aiming to determine head pose healthiness of computer users. The main contributions of this paper are: 1) Image Quality Assessment (IQA) based Stacked Autoencoder (referred to as IQASAE) which adjusts the value of learning rate based on the quality of images; 2) Head Pose Healthiness Prediction (HPHP) framework which leverages the proposed IQASAE algorithm in combination with image processing operations; 3) A set of features suitable for face analysis applications; 4) Ontology-driven semantic framework which enables further exploiting pose estimation results within applications in synergy with healthcare expert domain knowledge about pose healthiness. Our framework was evaluated on both offline (BIWI and AFLW) and online (our own, collected using Arduino) datasets. Furthermore, it was compared to several state-of-art methods, including Multi-Layer Perceptron (MLP), CART, Random Forest, Convolutional Neural Networks (CNN), Temporal Deep Learning Model (TDLM), hybrid CNN with Support Vector Machine (SVM), Quatnet and Trinet. According to the achieved experimental results, it reaches accuracy up to 79.63% outperforming all of them, except Quatnet and Trinet. However, the main advantages of IQASAE compared to state-of-art methods are: 1) it does not require selection of features, so the processing time is reduced, 2) utilizing angle between chin and mouth reduces training time for SAE, 3) leveraging vector-based feature set to create training data resulted in a significant improvement, especially in offline facial images. [ABSTRACT FROM AUTHOR]

As a specially designed tool and technique for the detection of image steganography, image steganalysis conceals information under the carriers for covert communications. Being developed on the BOSSbase dataset and released a decade ago, most of the Convolutional Neural Network (CNN) architectures for spatial image steganalysis fail to achieve satisfactory performance on new challenging datasets, i.e. ALASKA#2, which was released recently and is more complex yet consistent with the real scenarios. In this paper, we propose an enhanced residual network (ERANet) with self-attention ability, which utilizes a more complex residual method and a global self-attention technique, to alleviate the problem. Compared to the residual network that was widely used in the state-of-the-art, the enhanced residual network mathematically employed a more sophisticated way to extract more effective features in the images and hence it is suitable for more complex situations in the new dataset. Our proposed Enhanced Low-Level Feature Representation Module can be easily mounted on other CNNs in selecting the most representative features. Although it comes with a slightly extra computational cost, comprehensive experiments on the BOSSbase and ALASKA#2 datasets at various sizes have demonstrated the effectiveness of the proposed methodology. In short, ERANet provides an improvement of about 3.77% on average, compared to a few state-of-the-art CNNs. [ABSTRACT FROM AUTHOR]

Human Activity Recognition (HAR) is an emerging technology with applications in the surveillance, security, and healthcare sectors. Noninvasive HAR systems based on Wi-Fi Channel State Information (CSI) signals can be developed leveraging the rapid growth of ubiquitous Wi-Fi technologies and the correlation between CSI dynamics and body motions. In this paper, we propose Principal Component-based Wavelet Convolutional Neural Network (or PCWCNN) – a novel approach for HAR that offers robustness and efficiency for practical real-time applications. Our proposed method incorporates two efficient preprocessing algorithms – the Principal Component Analysis (PCA) and the Discrete Wavelet Transform (DWT). We employ an adaptive activity segmentation algorithm that is accurate and computationally light. Additionally, we used Wavelet CNN – a deep convolutional network analogous to the well-studied ResNet and DenseNet networks, for activity classification. We evaluate the performance of our proposed PCWCNN algorithm on both line-of-sight data and non-line-of-sight data. We empirically show that our proposed PCWCNN algorithm performs very well on these real datasets, outperforming relevant state-of-the-art approaches. [ABSTRACT FROM AUTHOR]

This paper investigates the problem of effective and robust fusion representation for foreground moving object detection. Many deep learning-based approaches pay attention to network architecture based on a single modality image, ignoring the complementary mechanism of cross-modal images or only considering single-frame prediction without temporal association. We tackle these problems by proposing a fusion representation learning method for the foreground moving object detection task, which consists of two major modules: the upstream fusion representation module (FRM) and the downstream foreground moving object detection module (FODM). Unlike traditional feature aggregate methods, the FRM module is a quality-aware and online learnable fusion module which can aggregate valuable features while rejecting the harmful information in the source images. Specifically, the FODM module is a siamese convolutional neural network to detect foreground moving objects by aggregating the time-sequence images generated by FRM. Moreover, a new aligned foreground moving object detection dataset of infrared and visible images is constructed to provide a new option for benchmark evaluation. Experimental results and comparisons with the state-of-the-art on three public datasets validate the effectiveness, robustness, and overall superiority of our method. [ABSTRACT FROM AUTHOR]

Residual learning is one of the most effective components in blind image denoising. It learns to estimate the noise instead of the clean image itself. A shortcoming of residual learning is that it cannot capture hierarchical features efficiently because there are no connections between layers to extract multi-level features. Therefore, residual dense network (RDN) adds dense connections among layers in the residual block to facilitate extraction of hierarchical features. Although RDN is a superior non-blind denoiser, it is only better than state-of-the-art methods in blind denoising when noise level is high. In this paper, we employ the concept of Wiener filters and add a non-residual network as a guidance for RDN. The non-residual network predicts the signal instead of the noise to assist RDN in achieving satisfactory performance across all noise levels. Then, a guidance network combines the outputs from both residual and non-residual networks to generate the final denoised output. Experimental results show that the new architecture performs better than state-of-the-art blind denoisers quantitatively and image quality of the new method is also better than existing methods. • Improve performance of residual dense network in blind image denoising by employing the concept of Wiener filters. • Use non-residual network as a guidance of residual dense network and assist in predicting the signal instead of the noise. • Multiplications of noisy input and outputs from networks are used for signal and noise power prediction. • A guidance network combines input, outputs from networks, and the above multiplications to generate final denoised output. • Experimental results show residual dense network with non-residual guidance performs better than state-of-the-art methods. [ABSTRACT FROM AUTHOR]

As an essential step in the early diagnosis of retinopathy, the blood vessels morphological attributes assist specialists to obtain pathological information efficiently. Most existing deep learning methods are based on U-shaped convolutional neural networks for the segmentation of blood vessels and have made substantial progress. However, the variance between vessel images remains challenging for segmentation algorithms, as demonstrated by poor cross-validation performance between different datasets. In this paper, a novel U-shaped deep convolutional network is proposed for retinal vessel segmentation, namely spiking neural P-type Dual-channel dilated convolutional network (SDDC-Net). We redesign the classical U-shaped convolution network based on the spiking neural P system computational mechanism for the first time. Distinct from the conventional convolutional neural network, SDDC-Net integrates the spiking neural P system convolutional neurons into the classic encoder-decoder architecture. We employ dilated convolution into an encoder, which improves both capabilities of perceiving more contexts and perceptual sensitivity of thin blood. We evaluate this model on three public datasets (DRIVE, STARE, CHASE_DB1), which indicates the more sensitive detection of thin vessels compared to most existing methods, showing higher sensitivity and F1 metrics. Compared to the baseline U-Net, our sensitivity, accuracy, and F1 score metrics on DRIVE dataset surpass by 10.66%, 1.73%, and 1.47% respectively. We also evaluate the effectiveness of spiking neural P system and dilated convolution in the ablation experiments, which demonstrates that accuracy increases with few drops in specificity. The cross-validation experiments show that our model has not only effective segmentation ability but also excellent generalization ability. [ABSTRACT FROM AUTHOR]

Deep convolutional neural networks (CNNs) often show inferior classification performance when training samples are seriously class-imbalanced. In this paper, we present a new Mixup-based regularization method, which loosens the constraint of only using synthetic samples to train the model. Specifically, we choose to cut and paste patches between the training images to generate samples, and restore corresponding minority image samples in the synthetic samples which contain a large proportion of minority image patches. At the same time, we restrict the location of patch generation, so as to intercept more image areas of training samples. By doing so, we increase the number of minority samples for the training set. The proposed method performs good generalization on minority classes by training the classifier to push the decision boundary toward majority classes. Besides, our approach integrates well with conventional rebalancing techniques such as re-weighting or re-sampling. We validate our method on the imbalanced datasets created by CIFAR-10, CIFAR-100, CINIC-10 and Tiny-ImageNet. The results demonstrate that our method is superior to some other Mixup-based regularization techniques (e.g., Mixup, Remix and Cutmix). For example, on the long-tailed version of CIFAR-10 (CIFAR-100) with an imbalance ratio of 100, our proposed method achieves a 4.90% (4.06%) improvement in classification accuracy compared with the classical Remix method. [ABSTRACT FROM AUTHOR]

Radar specific emitter identification (SEI) plays important roles in spectrum resource management and Internet of Things (IoT) security. This term refers to the identification of different individual devices of the same type by analyzing the unintentional modulation features caused by hardware imperfections (i.e., fingerprint features). Increasingly many researchers are using deep learning (DL)-based methods for SEI. However, DL-based methods still have many problems, including model generalization, recognition efficiency, and adaptability in open electromagnetic environments. This paper proposes a one-dimensional convolutional neural network (1D CNN)—called a selective kernel residual network (SKResNet)—capable of autonomously adjusting the receptive field, to improve the feature extraction capability. The transient part is cut efficiently from a radar signal sequence as the input of SKResNet using a universal preprocessing method. In addition, considering that most existing methods obtain incorrect results while identifying new emitters without any prior knowledge, SKResNet was improved using OpenMax—an open-set recognition method—to develop Open-SKResNet. As this network could identify unknown emitters effectively, generating reliable classification results of known emitters simultaneously, it could be of great importance and have broad application prospects. [ABSTRACT FROM AUTHOR]

With the help of some modern image generative techniques, it is possible to generate or manipulate image or video contents without introducing any obvious visual artifacts. If these manipulated images/videos are abused, it probably has a huge negative impact on society and individuals. Thus, deepfake detection has attracted considerable attention in recent years. Although the existing methods can achieve good detection performance on high-quality datasets, they are still far from satisfactory for low-quality dataset and cross-dataset evaluation. In this paper, therefore, we propose a new CNN-based method via multi-scale convolution and vision transformer for deepfake detection. In the proposed model, we design a multi-scale module with dilation convolution and depthwise separable convolution to capture more face details and tampering artifacts at different scales. Unlike the traditional classification module, furthermore, we employ a vision transformer to further learn the global information of face features for classification. Extensive experiments demonstrate that in most cases the proposed method achieves better detection results on both high-quality and low-quality datasets compared with related modern methods, and the cross-dataset generalization capabilities of the proposed method are good. In addition, many ablation experiments are provided to verify the rationality of the proposed network. [ABSTRACT FROM AUTHOR]

Understanding and analyzing radar work modes play a key role in electronic support measure system. Many classifiers, for example those based on convolutional neural network (CNN) and recurrent neural network (RNN), are available for recognizing radar work modes as well as emitter types from their waveform parameters. However, the performance of these methods may suffer significantly when confronting different types of signal degradation, e.g., measurement error, lost pulse and spurious pulse. To tackle this issue, we in this paper develop a Bayesian attention belief network (BABNet) based on Bayesian neural networks in which the probability distribution over weights can help to enhance the model robustness for corrupted data. In particular, we adopt pre-trained CNN as the Bayesian inference prior. This not only accelerates the convergence speed, but also avoids the training process getting stuck in bad local minima. Meanwhile, instead of using RNNs which are difficult to be implemented in parallel, the combination of padding operation and attention module in the proposed BABNet enables CNN, as the backbone, to process sequential data with variable length. Extensive experiments are conducted to demonstrate the recognition capability and robustness of the BABNet in different environments. • Proposed an attention-based Bayesian neural network for radar work mode recognition. • Introduced uncertainty on weights in neural networks to enhance detection robustness. • Suggested a learning prior for Bayesian variational inference. • Developed a convolution operation together with attention mechanism to replace RNNs. • Thorough experimental validation and comparison on different types of benchmark data. [ABSTRACT FROM AUTHOR]

In the realm of computer vision, Deep Convolutional Neural Networks (DCNNs) have demonstrated excellent performance. Video Processing, Object Detection, Image Segmentation, Image Classification, Speech Recognition and Natural Language Processing are some of the application areas of CNN. Object detection is the most crucial and challenging task of computer vision. It has numerous applications in the field of security, military, transportation and medical sciences. In this review, object detection and its different aspects have been covered in detail. With the gradual increase in the evolution of deep learning algorithms for detecting objects, a significant improvement in the performance of object detection models has been observed. However, this does not imply that the conventional object detection methods, which had been evolving for decades prior to the emergence of deep learning, had become outdated. There are some cases where conventional methods with global features are superior choice. This review paper starts with a quick overview of object detection followed by object detection frameworks, backbone convolutional neural network, and an overview of common datasets along with the evaluation metrics. Object detection problems and applications are also studied in detail. Some future research challenges in designing deep neural networks are discussed. Lastly, the performance of object detection models on PASCAL VOC and MS COCO datasets is compared and conclusions are drawn. [ABSTRACT FROM AUTHOR]

Acoustic source localization in the spherical harmonic domain with reverberation has hitherto not been investigated extensively. When the joint direction-of-arrival (DOA) estimation is treated as a classification task, it needs a lot of different classes and large-scale dataset, which is time-consuming and less cost-effective. Thus, a novel learning-based method is proposed in this paper, which separately estimates azimuths and elevations by two independent convolution neural networks (CNNs) with residual blocks. First, a mapping matrix is adopted to transform the spherical harmonic function into the Kronecker product of two Vandermonde vectors. Based on the special structure, features dependent on azimuth and elevation, respectively, are extracted as inputs to neural networks. This can divide the joint DOA estimation into two sub-tasks completed by two independent CNNs with residual blocks, which can be realized in parallel. Furthermore, the performance of the model learned by CNNs with residual blocks is better than ordinary CNNs. Simulations are conducted both on the simulated and real speech data for evaluating the performance of the proposed method. The results show that the method has higher accuracy. Moreover, the method is time-saving and effectively reduces the computational complexity. [ABSTRACT FROM AUTHOR]

Arbitrary shape scene text detection is still a challenging task in the computer vision community. Compared with anchor-based methods, segmentation-based methods have advantages in some respects, which detect text instances by generating accurate text components. In this paper, we propose a novel text components extraction network for arbitrary shape scene text detection. It can detect different text components through two parallel branches. In the first branch, the feature redistribution module (FRM) is proposed to extract text-related features and filter non-text information. Then, these text-related features are aggregated to generate text boundary maps. These text-related features are also fed to the second branch to progressively generate the text kernel maps. The second branch consists of a modified Transformer decoder with a multi-level progressive supervision strategy. It aims to capture spatial details and establish the dependencies between different text regions. In this way, our method can generate accurate text components without additional geometric calculations. Thanks to the powerful Transformer decoder and efficient differentiable binarization module, our method not only achieves advanced detection accuracy but also has a competitive inference speed. Specifically, with the ResNet-18 backbone, our method can run at 43.5 FPS and achieve an F-measure of 83.3% on the Total-Text dataset, 1.43 times faster than the latest state-of-the-art method. With the ResNet-50 backbone, our method achieves an F-measure of 85.2% and outperforms the latest state-of-the-art method by 0.2% on the Total-Text dataset. Code is available at: https://github.com/As-David/TransText. [ABSTRACT FROM AUTHOR]

As moving objects always draw more attention of human eyes, the temporal motion information is always exploited complementarily with spatial information to detect salient objects in videos. Although efficient tools such as optical flow have been proposed to extract temporal motion information, it often encounters difficulties when used for saliency detection due to the movement of camera or the partial movement of salient objects. In this paper, we investigate the complementary roles of spatial and temporal information and propose a novel dynamic spatiotemporal network (DS-Net) for more effective fusion of spatiotemporal information. We construct a symmetric two-bypass network to explicitly extract spatial and temporal features. A dynamic weight generator (DWG) is designed to automatically learn the reliability of corresponding saliency branch. And a top-down cross attentive aggregation (CAA) procedure is designed to facilitate dynamic complementary aggregation of spatiotemporal features. Finally, the features are modified by spatial attention with the guidance of coarse saliency map and then go through decoder part for final saliency map. Experimental results on five benchmarks VOS, DAVIS, FBMS, SegV2, and ViSal demonstrate that the proposed method achieves superior performance than state-of-the-art algorithms. The source code is available at https://github.com/TJUMMG/DS-Net. [ABSTRACT FROM AUTHOR]

Automatic modulation recognition (AMR) plays an essential role in wireless communication systems. Our paper proposes a novel multi-stream neural network (MSNN) to extract the features in parallel from the amplitude, phase, frequency, and raw data of the modulated signal. The framework integrates convolutional neural networks (CNN) and bidirectional gated recurrent units (Bi-GRU) to extract features more effectively from the spatial and temporal characteristics with the assistance of two different attention mechanisms, convolutional block attention module (CBAM) and multi-head self-attention (MHSA). Simulation experiments show that the performance of our proposed algorithm is better than that of other state-of-the-art (SOTA) recognition algorithms on four widely used DeepSig datasets. The recognition accuracy of our proposed model exceeds 99% (18 dB), 93% (16 dB), 95% (20 dB), and 97% (20 dB) respectively on the datasets RML2016.04C, RML2016.10A, RML2016.10B, and RML2018.01A. [ABSTRACT FROM AUTHOR]

With the rapid development of convolutional neural networks (CNNs), the performance of human pose estimation has been significantly improved. However, state-of-the-art methods still face specific challenges, such as occluded keypoints and nearby persons. Unlike the classical network improvement strategy, our approach obtains adaptive data to obtain more accurate keypoints through data augmentation. In this paper, we propose Adaptive Data Augmentation Network (ADA-Net), which brings more adaptive training data by adding occlusion and interleaving information to the original image. First, we introduce Active Transmission Network (ATNet), which actively learns the transformation matrix from the original image and uses the matrix to synthesize new images for training. Second, we adopt an adversarial training strategy combined with the ATNet that allows us to capture more challenging cases. Extensive experiments show that our approach achieves comparable or even better results than most state-of-the-art methods. In particular, our ADA-Net outperforms High-Resolution Network (HRNet) by 1.1 and 0.7 points on the COCO test-dev set and MPII validation set, respectively. • The diversity of training samples will enhance the performance of human pose estimation. • The appropriate number of pasted samples on the original image can affect network performance. • Pasting the complete person is beneficial in enhancing the original image information. • The generative adversarial network brings the synthetic image closer to real challenging cases. [ABSTRACT FROM AUTHOR]

Intersection is an important component in urban traffic networks, as it connects pedestrian flows and vehicles between the network links. Intersection classification plays an active role in mitigating traffic congestion, increasing the level of road safety and traffic efficiency. Various intersection classification methods are adopted to classify the intersection point in the road segment, but detecting the accurate location results from a complex task in an automatic driving system. Hence, this paper proposes a Rider Border Collie Optimization-based Deep Convolutional Neural Network (RBorderNet) for road scene segmentation intersection classification. The RBCO model is modeled by combining the Rider Optimization Algorithm (ROA) with Border Collie Optimization (BCO). Here, Fresnel Transform (FrT) is employed to detect the keyframes from the video based on the angular distance. DCNN classifier is used for classifying the intersection of road segments by considering the optimal region extracted from the segmented frames. The training of the DCNN classifier is accomplished by the RBCO algorithm. The developed RBCO-based DCNN achieved higher performance with the metrics, like accuracy, training error, precision, and recall is 94.51%, 5.49%, 96.43%, and 95.29%, respectively, by considering vehicles and traffic scenarios. [ABSTRACT FROM AUTHOR]

Cascaded channels have been considered in several physical multipath propagation scenarios. However they are subject to phenomena such as multipath scattering, time dispersion and Doppler shift between the different links, which impose great challenges in relation to the channel estimation processing function in the receiver. In this paper we propose to tackle the problem of cascaded channels estimation in the fifth-generation and beyond (5G+) systems using convolutional neural networks (CNNs), without forward error correction (FEC) codes. The results show that the CNN-based framework reaches very close to perfect (theoretical) channel estimation levels, in terms of bit error rate (BER) values, and outperforms the least square (LS) practical estimation, measured in mean squared error (MSE). The benefits of CNN-based wireless cascaded channels estimation are particularly relevant for increasing number of links and modulation order. These findings are further confirmed at the CNN implementation stage on a field programmable gate array (FPGA) platform for a number of realistic quantization scenarios. [ABSTRACT FROM AUTHOR]

Nowadays it is common to see image forgeries on almost every media in both professional and personal contexts. From visual retouches to deliberate fake scenes, the technology used to create image forgeries gets easier to use for all users. At the same time, different techniques to assess the authenticity of the content of an image have appeared by taking advantage of the deep learning paradigm. In this paper we propose two initialization approaches for Convolutional Neural Networks (CNNs) used for the detection of image manipulation operations. We focus on the variance stability for the output of a convolutional filter in CNN. Our first proposal is a scaling approach for first-layer convolutional kernels which can cope well with filters generated by different algorithms. Our second proposal is a random high-pass filter initialization approach for CNN's first convolutional layer. The first proposal explicitly computes simple statistical properties of the input signal, while the second approach incorporates the consideration of input statistics in the filter derivation without the need of carrying out explicit computation on the input. Experimental results show the utility of both approaches with improved performance in different image manipulation detection problems and on different CNN architectures. • Initializations of convolutional neural networks for image forensics. • Data-dependent and data-independent initialization approaches. • Mathematical derivations by considering natural image statistics. • Performance improvement of image manipulation detection. [ABSTRACT FROM AUTHOR]

This paper proposes a combined Convolutional Neural Network (CNN) scheme appropriate for automatic modulation recognition. The suggested network includes a deep network involving multiscale convolutions with an attention module for recognizing 24 modulation modes and a shallow one network for recognizing two modulation modes, aiming to reduce the misjudgment caused by the deeper network. We evaluate our proposed network on the DeepSig dataset achieving an accuracy of 98.7% that exceeds the performance of current state-of-the-art (SOTA) methods exploiting this dataset. The network achieved almost 100% accuracy on modulation modes of the order less than or equal to 64 and under high SNR. The accuracy attained by our method on high-order modulation modes exceeds the one of ResNet, which currently is the SOTA method. [ABSTRACT FROM AUTHOR]

Blind identification of channel codes is becoming increasingly important in signal interception and intelligent communication systems. However, most existing channel codes recognition algorithms extract features manually, which makes them highly demanding in real-world application. Thus, efficiently identifying channel codes is difficult using present technologies. This paper presents a deep residual network-based deep learning (DL) approach on the blind identification of convolutional code parameters for a given soft-decision sequence. The proposed method can blindly identify the convolutional codes without the need for the prior information about its coding parameters, and it achieves over 88% of recognition accuracy for 17 forms of convolutional codes when SNR exceeds or equals zero. Furthermore, we investigate factors affecting the accuracy of channel codes recognition including input length, model depth and data type. A comparison of the recognition accuracy between the proposed algorithm, log-likelihood ratio (LLR)-based traditional blind identification algorithm, and DL-based algorithm are then made. Experiment results show that deep residual network-based approaches could provide significant improvements over the traditional algorithm or existing DL-based algorithms in the blind identification of convolutional codes. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel differentiable activation function for convolutional neural networks. This function is proposed using an orthogonal Legendre polynomial based linear-in-parameter model. First, we present mathematical modeling for max pooling operation in convolutional neural networks and derive the relationship between max pooling and average pooling. The representation of max pooling using signal processing elements is presented. Further, we study approximations of various nonlinearities using the proposed linear-in-parameter model. The proposed activation function can be used as an alternative in place of existing activation functions in convolutional neural networks. Finally, the effectiveness of the proposed activation function is shown with empirical evaluations on benchmark image classification datasets using convolutional neural networks. [ABSTRACT FROM AUTHOR]

Three main requirements of a successful application of deep learning are the network architecture, a large enough training dataset, and a good optimization algorithm. In this paper we mainly focus on the optimization part. We propose a training algorithm for convolutional neural networks which makes use of both first and second order derivatives for training different layers. We utilize an approximate second order algorithm for the classification layer while we train the rest of the network with the conventional approach which is backpropagation with first order derivatives. We show that this approach helps us achieve a higher classification accuracy with a much smaller number of training iterations compared to training the whole network with gradient descent based algorithms. Moreover, although second order optimization is generally costlier, we show that the proposed approach is trained faster not only in terms of the number of iterations but also training duration. We also present the integration of CNNs with a probabilistic spatial model and apply this to the land cover classification problem in hyperspectral images. The results show that the algorithm allows us to achieve superior results with a simple network even with limited training data compared to existing approaches. [ABSTRACT FROM AUTHOR]

An error correcting code type recognition technique based on a deep learning approach is proposed in this paper. This problem could be addressed in the context of non-cooperative communications or adaptive coding and modulation. Inspired by text classification, we proposed a convolutional neural network (CNN) model improved by embedding and block mechanism to classify the linear block code, convolutional code, and turbo code with the only knowledge of the noisy information streams. It achieves higher recognition performance than the algorithms which are based on traditional deep learning and rank calculation. Further results show that the performance is greatly affected by block length and the dimension of the embedding layer. In a nutshell, the CNN with block mechanism and embedding is a promising feature extraction and classification technique, and it is suitable for the recognition of different kinds of communication signals. [ABSTRACT FROM AUTHOR]

• A deep-learning method for classification of the audio codecs is proposed. • The effect of variable bit-rates in the classification of fragments is analyzed. • Longest common substrings and subsequences can effectively classify audio codecs. • A large set of speech codecs with 20 codecs and 96 bit-rates is considered. A large portion of the Internet bandwidth is used for transmission of multimedia such as audio data. As the file sizes are usually much bigger than the maximum network packet size, the audio data are segmented into fragments. For eavesdropping or network surveillance purposes, the first step of a sniffer may be to determine the codec by which a fragment is generated. This problem is usually modeled as a multi-class classification problem. The basic methods for determining the codec type of each fragment rely on the metadata in the corresponding file header. However, in a non-cooperative context, the whole file is not available. Statistical features extracted from the fragments are generally used for solving this problem by employing machine learning methods. In this paper, we present an end to end scalable deep learning approach for the classification of audio codecs with variable bit-rates. This method is based on an efficient variant of convolutional neural networks, which learns hidden layer representations to encode input vectors. Moreover, by integrating the statistical features and Cascade Deep Forest method, we jointly optimize classification labels assignment and learn the features that are suitable for classification. In several experiments on a dataset of speech codecs, we demonstrate that our approach outperforms the state-of-the-art methods by a significant margin. [ABSTRACT FROM AUTHOR]

Specific emitter identification (SEI) is an important problem in the field of electronic intelligence. There are two major limitations in most existing SEI methods: First, the features should be artificially extracted, which requires specialized expertise; Second, various features are not merged effectively to improve performance. In this paper, an automatic multi-feature extraction and fusion method based on deep ensemble learning is proposed for SEI. This method extracts and fuses multiple features via data-driven strategies using convolutional neural networks (CNN). No additional constraints are required on the type and number of the features to be fused. Three typical characteristics of radar pulse signals, including amplitude, phase and spectrum asymmetry, are taken as example features for method description and performance verification. Experiment results demonstrate that, the proposed method performs well in automatic feature extraction, and it can significantly improve the SEI performance via multi-feature fusion. [ABSTRACT FROM AUTHOR]

SAR images are a high-resolution map of surface target areas and terrain in the range and the cross-range dimension. Usually, moving targets appear as defocused and spatially displaced objects superimposed on the SAR map. In this paper, a new moving target refocusing imaging method based on Range Doppler (RD) Algorithm and convolutional neural network is proposed. Firstly, Range Doppler (RD) Algorithm is used to preprocess the echo data to obtain the blurring SAR image as the input data of convolution neural network. Secondly, according to SAR imaging characteristics, the original U-net structure is improved to extract the structure information of input data and provide good target reconstruction. The trained network can focus moving targets in case of different azimuth velocities. Finally, the experimental results prove the effectiveness of the proposed method in radar image focusing. [ABSTRACT FROM AUTHOR]

In the computer vision, object detection has always been considered one of the most challenging issues because it requires classifying and locating objects in the same scene. Many object detection approaches were recently proposed based on deep convolutional neural networks (DCNNs), which have been demonstrated to achieve outstanding object detection performance compared to other approaches. However, the supervised training of DCNNs mostly uses gradient-based optimization criteria, in which all parameters of hidden layers require multiple iterations, and often faces some problems such as local minima, intensive human intervention, time-consuming, etc. In this paper, we propose a new method called Faster-YOLO, which is able to perform real-time object detection. The deep random kernel convolutional extreme learning machine (DRKCELM) and double hidden layer extreme learning machine auto-encoder (DLELM-AE) joint network is used as a feature extractor for object detection, which integrating the advantages of ELM-LRF and ELM-AE. It takes the raw images directly as input and thus is suitable for the different datasets. In addition, most connection weights are randomly generated, so there are few parameter settings and training speed is faster. The experiment results on Pascal VOC dataset show that Faster-YOLO improves the detection accuracy effectively by 1.1 percentage points compared to the original YOLOv2, and an average 2X speedup compared to YOLOv3. [ABSTRACT FROM AUTHOR]