Search

Dual-layer displays (DLDs) can achieve high dynamic range by reducing light leakage via bonding two LC panels. However, the physical distance between the two LC panels introduces a parallax error when the display is viewed off-axis. This paper proposes a convolutional neural network-based image splitting algorithm for improving the display quality of a DLD. In order to improve processing speed, the network structure is designed relatively simply. The training dataset is established mainly based on TV signals. Different from the common image evaluation metrics based on PSNR and SSIM, the paper proposes to apply learned perceptual image patch similarity (LPIPS) to define the loss function for DLDs. The simulation and experiment results show that compared with the existing method, the proposed method has better performance in reducing parallax error and improving image details. The processing time can be shortened by 2.7 times. The proposed CNN-based method can greatly reduce the processing time while maintaining high display quality, which presents high practicability. [ABSTRACT FROM AUTHOR]

This paper is concerned with the autonomous effective collision avoidance strategy for multiple unmanned aerial vehicles (multi-UAV) in limited airspace under the framework of proximal policy optimization (PPO) algorithm. An end-to-end deep reinforcement learning (DRL) control strategy and a potential-based reward function are designed. Next, the CNN-LSTM (CL) fusion network is constructed by fusing the convolutional neural network (CNN) and the long short-term memory network (LSTM), which realizes the feature interaction among the information of multi-UAV. Then, a generalized integral compensator (GIC) is introduced into the actor-critic structure, and the CLPPO-GIC algorithm is proposed by combining CL and GIC. Finally, we validate the learned policy in various simulation environments by performance evaluation. The simulation results show that the introduction of the LSTM network and GIC can further improve the efficiency of collision avoidance, and the robustness and accuracy of the algorithm are verified in different environments. [ABSTRACT FROM AUTHOR]

• The gated recurrent unit algorithm with fused attention mechanism is a target positioning method in a non-line-of-sight error environment. • The introduction of the attention mechanism can analyze features, select important features, and better handle non-line-of-sight errors. • Introducing the dropout layer during training reduces the risk of overfitting during training. • Through simulation and actual measurement experiments, the effectiveness of the algorithm in this paper is verified by comparing it with other famous algorithms. Amidst the increasing reliance on indoor positioning in areas such as large shopping malls, airports, and train stations, the precision of ultra-wideband (UWB) technology stands out for its centimeter-level accuracy. However, non-line-of-sight (NLOS) errors caused by people's movements significantly reduce indoor positioning accuracy. In response, this paper presents an innovative blend of gated recurrent unit algorithm with fused attention mechanism (GRU_Attention). This algorithm adeptly captures the dynamism in UWB signals and intelligently focuses on pivotal features to actualize accurate three-dimensional positioning while effectively reducing NLOS error. It has been demonstrated that GRU_Attention algorithm can reach 5.6 cm accuracy level through experiments, which improved by 74.47 %, 87.44 %, 14.46 %, 43.82 %, and 29.07 % respectively compared with the backpropagation (BP), least square (LS), GRU, convolutional neural network (CNN), and fused attention mechanism convolutional neural network (CNN_Attention) algorithms. This algorithm charts a new pathway to overcome NLOS errors in indoor personnel movement scenes. [ABSTRACT FROM AUTHOR]

Intrusion Detection System (IDS) can ensure the network security by identifying network intrusions according to the abnormal traffic data. However, the intrusion detection data has the problem of high dimensionality and changes with network and attack environments, which leads to the poor performance and poor portability of intrusion detection algorithms. Therefore, this paper proposes an intrusion detection algorithm based on joint symmetric uncertainty and hyperparameter optimized fusion neural network. Firstly, a feature selection method based on symmetric uncertainty and approximate Markov blanket is proposed, which fully considers the correlation and redundancy of features, and also the correlation between combined features and the class label, so as to reduce the data dimensionality. Secondly, the CNN-LSTM classifier fused with Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) is used to extract the spatial features and temporal features to improve the classification performance. Finally, the Particle Swarm Optimization (PSO) algorithm is improved and used to automatically optimize the hyperparameters of the classifier, so that the classifier can be applied to different intrusion detection datasets with better generalization ability and portability. Experiments have verified the effectiveness and superiority of the proposed algorithm on multiple evaluation indicators. • An effective algorithm for intrusion detection is proposed in this paper. • Feature selection is based on symmetric uncertainty and approximate Markov blanket. • A fusion neural network is constructed to extract the spatial and temporal features. • The PSO algorithm is improved to automatically optimize the hyperparameters. [ABSTRACT FROM AUTHOR]

• An edge detection algorithm for thermal infrared weak texture images. • Construct a high-quality thermal infrared image edge dataset RI-BSDS500 for training to improve the training convergence speed and generalization ability of the model. • Different thermal infrared image datasets are tested under different algorithms, and the results are analyzed from the perspectives of objective parameters and subjective effects, which proves that the proposed algorithm in this paper is more effective on thermal infrared images. In order to solve the problems such as difficulties in extracting image edge information caused by low contrast and blurred edges of thermal infrared images, a thermal infrared image edge detection algorithm PiDiNet-TIR constructed based on the PiDiNet model is proposed. On the basis of the visible light image dataset BSDS500, the visible light image is grayed out, noise is added, and contrast is reduced to construct the thermal infrared image edge dataset I R-BSDS500, which provides image data for model training and testing. Deeply study the theory of PiDiNet model, add rich convolution mechanism to its backbone network, improve the activation function, pooling layer and up-sampling module, study the SimAM attention mechanism and add it to the side network. Using the open-source thermal infrared image dataset and homemade thermal infrared images for testing, we compare and analyze the mean square error (MSE), structural similarity of image (SSIM), feature similarity of image (FSIM), and Frames Per Second (FPS)of different detection algorithms, such as HED, RCF, DexiNet, PiDiNet, and the algorithm proposed in this paper. The results of image evaluation parameters prove that the edge detection algorithm proposed in this paper has high feature extraction efficiency and feature expression ability, and the algorithm detection performance is higher than other algorithms. [ABSTRACT FROM AUTHOR]

• The accuracy is greatly improved after replacing SCS module with the CSC module in Shift-GCN. • Inspired by shift CNNs, we replace shift space convolution in Shift-GCN with a sparse shift and named SparseShift-GCN. • The accuracy of algorithm was further improved after introduction of OHEM into SparseShift-GCN. Skeleton-based action recognition is widely used due to its advantages of lightweight and strong anti-interference. Recently, graph convolutional networks (GCNs) have been applied to action recognition and have made breakthrough progress. The shift convolution operator can effectively replace the spatial convolution and greatly reduce the computational complexity of the algorithm. This article first applies the Conv-Shift-Conv (CSC) module and the Shift-Conv-Shift-Conv (SC2) module to replace the Shift-Conv-Shift (SCS) module in spatial graph convolution of Shift-GCN respectively. This design can reorder the shifted channels more effectively. The experimental results show that the CSC module has the best effect and effectively improves accuracy of model. After that, this article proposes to replace the shift module in the original Shift-GCN with a sparse shift module and named SparseShift-GCN. This structure can reduce the redundancy of features, prevent overfitting and improve the generality of the model. Based on the improvement in the previous step, better results have been achieved. Finally, this paper uses OHEM Loss and Weighted Loss to carefully design the loss function of the model and introduces it into the model proposed in this paper. Experimental results show that OHEM Loss further improves the accuracy of algorithm. After a series of improvements, our proposed model has improved the accuracy of 4 different streams to varying degrees, which improves the overall performance of the network. [ABSTRACT FROM AUTHOR]

• Obtained a lightweight model by designing a streamlined network structure. • Improved the average detection accuracy of the model while keeping it lightweight. • This lightweight model performs exceptionally well on embedded edge devices. Infrared-based road scene object detection algorithms often face issues with excessive parameters and computational demands, making them incompatible with edge devices having constrained computational capabilities. This paper introduces an enhanced lightweight infrared-based road object detection algorithm based on YOLOv8n. Firstly, a streamlined network architecture is devised by merging YOLOv8n's C2f module with PConv, creating a lighter module and reducing the neural network's downsampling rate of infrared images. This strategy reduces redundant computations and memory access, preventing the loss of fine details in infrared images caused by deep convolutional neural networks. Additionally, the model's accuracy in detecting infrared targets is significantly enhanced through the integration of the coordinate attention mechanism. Finally, replacing CIoU with Wise-IoU for bounding box regression in YOLOv8n accelerates the model's convergence. Empirical findings indicate that in contrast to the YOLOv8n algorithm, the optimized model showcases a 34.17 % reduction in model size, a 40.35 % decrease in parameters, and a 4.8 % increase in average detection accuracy. This enhanced algorithm not only achieves a lightweight profile but also delivers superior performance on embedded edge devices. [ABSTRACT FROM AUTHOR]

As the deployment of artificial intelligence (AI) algorithms at edge devices becomes increasingly prevalent, enhancing the robustness and reliability of autonomous AI-based perception and decision systems is becoming as relevant as precision and performance, especially in applications areas considered safety-critical such as autonomous driving and aerospace. This paper delves into the robustness assessment in embedded Deep Neural Networks (DNNs), particularly focusing on the impact of parameter perturbations produced by single event upsets (SEUs) on convolutional neural networks (CNN) for image semantic segmentation. By scrutinizing the layer-by-layer and bit-by-bit sensitivity of various encoder–decoder models to soft errors, this study thoroughly investigates the vulnerability of segmentation DNNs to SEUs and evaluates the consequences of techniques like model pruning and parameter quantization on the robustness of compressed models aimed at embedded implementations. The findings offer valuable insights into the mechanisms underlying SEU-induced failures that allow for evaluating the robustness of DNNs once trained in advance. Moreover, based on the collected data, we propose a set of practical lightweight error mitigation techniques with no memory or computational cost suitable for resource-constrained deployments. The code used to perform the fault injection (FI) campaign is available at https://github.com/jonGuti13/TensorFI2 , while the code to implement proposed techniques is available at https://github.com/jonGuti13/parameterProtection. [ABSTRACT FROM AUTHOR]

Raman spectroscopy is a general and non-destructive detection technique that can obtain detailed information of the chemical structure of materials. In the past, when using chemometric algorithms to analyze the Raman spectra of mixtures, the challenges of complex spectral overlap and noise often limited the accurate identification of components. The emergence of deep learning has introduced a novel approach to qualitative analysis of mixed Raman spectra. In this paper, we propose a deep learning-based Raman spectroscopy qualitative analysis algorithm (RST) by borrowing the ideas of convolutional neural network and Transformer. By transforming the Raman spectrum into 64 word vectors, the contribution weights of each word vector to the components are obtained. For the 75 spectral data used for validation, the positive identification rate can reach 100.00 %, the recall rate can reach 99.3 %, the average identification score can reach 9.51, and it is applicable to the fields of Raman and surface-enhanced Raman spectroscopy. Furthermore, compared with traditional CNN models, RST has excellent accuracy and robustness in identifying components in complex mixtures. The model's interpretability has been enhanced, aiding in a deeper understanding of spectroscopic learning patterns for future analysis of more complex mixtures. [Display omitted] • Obtaining contribution weights for Raman spectra in different bands, the network exhibits strong interpretability. • Applicable in both Raman and surface-enhanced Raman spectroscopy fields. • Strong generalization ability, resistant to interference, and minimally affected by baseline noise. [ABSTRACT FROM AUTHOR]

With the increasing maturity of convolution neural network (CNN) technology, the image super-resolution reconstruction (SR) method based on CNN is booming and has achieved many remarkable results. Undoubtedly, SR has become the mainstream direction of image reconstruction technology. However, most of the existing SR methods improve the reconstruction performance by increasing the depth of networks, which also increases the number of parameters, number of network computations, and difficulty of training network. To solve the performance complexity dilemma in SR, this paper proposes a network called a multi-scale dense recursive fusion network (MSDRFN). The network is composed of three parts: initial feature extraction module, multi-scale dense fusion group module and recursive reconstruction module. In detail, rough features are first extracted through a shallow feature extraction module, and then are inputted into multi-scale dense fusion blocks (MSDFBs) group. Each MSDFB makes full use of image features in convolution kernels of different sizes to obtain different hierarchical features, and further these output features are inputted into the channel attention mechanism to learn their corresponding weights. All MSDFBs outputs will be restored to high resolution images via the recursive reconstruction module. In addition, the network supplements the information loss with residual learning, which is embodied in one long-jump connection and several short-jump connections. The proposed network is mainly trained in the Pytorch deep learning framework. In comparison experiments on benchmark datasets, the proposed method outperformed the most advanced convolutional methods. [ABSTRACT FROM AUTHOR]

The technology for Visual Odometry (VO) that estimates the position and orientation of the moving object through analyzing the image sequences captured by on-board cameras, has been well investigated with the rising interest in autonomous driving. This paper studies monocular VO from the perspective of Deep Learning (DL). Unlike most current learning-based methods, our approach, called DeepAVO, is established on the intuition that features contribute discriminately to different motion patterns. Specifically, we present a novel four-branch network to learn the rotation and translation by leveraging Convolutional Neural Networks (CNNs) to focus on different quadrants of optical flow input. To enhance the ability of feature selection, we further introduce an effective channel-spatial attention mechanism to force each branch to explicitly distill related information for specific Frame to Frame (F2F) motion estimation. Experiments on various datasets involving outdoor driving and indoor walking scenarios show that the proposed DeepAVO outperforms the state-of-the-art monocular methods by a large margin, demonstrating competitive performance to the stereo VO algorithm and verifying promising potential for generalization. [ABSTRACT FROM AUTHOR]

In this paper, we introduce convolutional proximal neural networks (cPNNs), which are by construction averaged operators. For filters with full length, we propose a stochastic gradient descent algorithm on a submanifold of the Stiefel manifold to train cPNNs. In case of filters with limited length, we design algorithms for minimizing functionals that approximate the orthogonality constraints imposed on the operators by penalizing the least squares distance to the identity operator. Then, we investigate how scaled cPNNs with a prescribed Lipschitz constant can be used for denoising signals and images, where the achieved quality depends on the Lipschitz constant. Finally, we apply cPNN based denoisers within a Plug-and-Play framework and provide convergence results for the corresponding PnP forward-backward splitting algorithm based on an oracle construction. [ABSTRACT FROM AUTHOR]

This paper illustrates the efficacy of an artificial intelligence (AI) (a convolutional neural network, based on the U-Net), for the burn-depth assessment using semantic segmentation of polarized high-performance light camera images of burn wounds. The proposed method is evaluated for paediatric scald injuries to differentiate four burn wound depths: superficial partial-thickness (healing in 0-7 days), superficial to intermediate partial-thickness (healing in 8-13 days), intermediate to deep partial-thickness (healing in 14-20 days), deep partial-thickness (healing after 21 days) and full-thickness burns, based on observed healing time. In total 100 burn images were acquired. Seventeen images contained all 4 burn depths and were used to train the network. Leave-one-out cross-validation reports were generated and an accuracy and dice coefficient average of almost 97% was then obtained. After that, the remaining 83 burn-wound images were evaluated using the different network during the cross-validation, achieving an accuracy and dice coefficient, both on average 92%. This technique offers an interesting new automated alternative for clinical decision support to assess and localize burn-depths in 2D digital images. Further training and improvement of the underlying algorithm by e.g., more images, seems feasible and thus promising for the future. [ABSTRACT FROM AUTHOR]

Normalized least mean squares (NLMS) and affine projection algorithm (APA) are two successful algorithms that improve the stability of least mean squares (LMS) by reducing the necessity to change the learning rate during the training process. In this paper, we extend them to multi-layer neural networks. We first prove that it is possible to consider a multi-layer neural network as a stack of adaptive filters. It opens the door to bring successful algorithms from adaptive filters to neural networks. We additionally introduce a more comprehensible interpretation than the complicated geometric interpretation in APA for a single fully-connected (FC) layer that can easily be generalized, for instance, to convolutional neural networks and mini-batch training. With this new viewpoint, we introduce a more robust algorithm by predicting and then amending the adverse consequences of some actions that take place in mini-batch backpropagation (BP), even before they happen. The proposed method is a modification to the BP that can be used alongside stochastic gradient descent (SGD) and its momentum variants like Adam and Nesterov. Our experiments show the usefulness of the proposed method in the training of deep neural networks. It is less sensitive to hyper-parameters and needs less intervention during the training process. Besides, it usually converges more smoothly and in fewer iterations. Such predictable behavior helps it to get tuned easier, be resilient during the training, and reduce or eliminate its reliance on other techniques such as momentum. [ABSTRACT FROM AUTHOR]

Lossy image compression is likely to produce blurred images which leads to erroneous image-level understanding. The human visual system (HVS) is focused on the area of interest present in the image. Motivated by this fact, we propose a compression-decompression algorithm focusing on the important content present in different parts of the image. Using the concept of the Guided Grad-CAM (Gradient-weighted Class Activation Mapping) technique to produce heat maps with the help of a heat map generator trained on ResNet-50, a saliency-guided encoding–decoding algorithm is developed. A wider multi-scale saliency guided convolutional neural network (MsSG-CNN) is designed, in which the notion of convolution with different size filters helps to obtain unique but different features. The feature extraction followed by multilevel fusion of features helps deep neural network (DNN) to capture contextual information and obtain high quality, good resolution, visually pleasing images with fine details. The proposed algorithm is tested on the Kodak benchmark dataset, CLIC 2019 challenging dataset, and FDDB facial images dataset. At low bit rates, the MS-SSIM of the proposed algorithm is found to be superior to JPEG, JPEG2000, BPG, WebP, and Minnen's approaches with approximately up to 60%, 24.80%, 11.43%, 23.08% & 75% gains respectively, which is quite a significant improvement, when tested on the Kodak dataset. Similarly, at high bit rates, the improvement in MS-SSIM is approximately up to 41.67%, 37.30%, 23.90% 34.21% & 13.33% when compared with JPEG, JPEG2000, BPG, WebP, and Minnen's approaches respectively. The improvement in PSNR at low and high bit rates is approximately up to 11.32%, 5%, 5.26% and 5.6%, 10.29%, 8.7% as compared to JPEG, Balle's, and Lee's algorithms respectively. The PSNR-HVS has been improved by approximately up to 27.27%, 19.15%, and 28.33%, 28% as compared to JPEG and Toderici's algorithms respectively at low and high bit-rates. A similar type of improvement is obtained with FDDB and CLIC 2019 datasets also, which is discussed in the paper. [ABSTRACT FROM AUTHOR]

In vitro microinjection is one of the main approaches to deliver the specimens to the target. However, conventional microinjection is mostly operated by very experienced technicians and the successful rate is relatively low. Moreover, the manual injection process is also very time consuming and tiring because extra attention needs to be paid for precise operation. As a result, the workers suffer from fatigue and the efficiency remains low. To improve the efficiency, researchers have developed automated microinjection systems to free the technicians from this laborious work. Zebrafish has been more popular than ever for the last decade due to its genetic advantages. It is claimed that the gene of Zebrafish highly resembles to human, which results in the fact that more biological and medical practice has been carried out with Zebrafish larvae or embryos. This paper proposes an innovation convolutional neural network based visual processing strategy for biological sample manipulation and microinjection systems. The proposed approach allows high throughput injection with considerable injecting precision. The strategy is to firstly locate a number of biological samples in the visual field, then the proposed visual algorithm identifies the number of the samples and detects the orientation of each sample. After which the samples are rotated one after another by a rotation plate to the desired angle for injection. An optimized rotation strategy is performed to achieve minimum total rotation angle. A path free of collision is also generated by A-star algorithm for the holding pipette. The experiment results show significant improvement in orientation efficiency and the accuracy of multi-sample rotation is enhanced to some extent. [ABSTRACT FROM AUTHOR]

Lossy image compression is likely to produce blurred images which leads to erroneous image-level understanding. The human visual system (HVS) is focused on the area of interest present in the image. Motivated by this fact, we propose a compression-decompression algorithm focusing on the important content present in different parts of the image. Using the concept of the Guided Grad-CAM (Gradient-weighted Class Activation Mapping) technique to produce heat maps with the help of a heat map generator trained on ResNet-50, a saliency-guided encoding–decoding algorithm is developed. A wider multi-scale saliency guided convolutional neural network (MsSG-CNN) is designed, in which the notion of convolution with different size filters helps to obtain unique but different features. The feature extraction followed by multilevel fusion of features helps deep neural network (DNN) to capture contextual information and obtain high quality, good resolution, visually pleasing images with fine details. The proposed algorithm is tested on the Kodak benchmark dataset, CLIC 2019 challenging dataset, and FDDB facial images dataset. At low bit rates, the MS-SSIM of the proposed algorithm is found to be superior to JPEG, JPEG2000, BPG, WebP, and Minnen's approaches with approximately up to 60%, 24.80%, 11.43%, 23.08% & 75% gains respectively, which is quite a significant improvement, when tested on the Kodak dataset. Similarly, at high bit rates, the improvement in MS-SSIM is approximately up to 41.67%, 37.30%, 23.90% 34.21% & 13.33% when compared with JPEG, JPEG2000, BPG, WebP, and Minnen's approaches respectively. The improvement in PSNR at low and high bit rates is approximately up to 11.32%, 5%, 5.26% and 5.6%, 10.29%, 8.7% as compared to JPEG, Balle's, and Lee's algorithms respectively. The PSNR-HVS has been improved by approximately up to 27.27%, 19.15%, and 28.33%, 28% as compared to JPEG and Toderici's algorithms respectively at low and high bit-rates. A similar type of improvement is obtained with FDDB and CLIC 2019 datasets also, which is discussed in the paper. [ABSTRACT FROM AUTHOR]

• We introduced an adaptive layer into the full connection layer of a deep convolution neural network. • The goal of the adaptive layer is to minimize the intra-class distance and maximize the inter-class distance. • The results revealed that the performance of the proposed algorithm was better than that of other algorithms. The nonstationarity problem of EEG is very serious, especially for spontaneous signals, which leads to the poor effect of machine learning related to spontaneous signals, especially in related tasks across time, which correspondingly limits the practical use of brain-computer interface (BCI). In this paper, we proposed a new transfer learning algorithm, which can utilize the labeled motor imagery (MI) EEG data at the previous time to achieve better classification accuracies for a small number of labeled EEG signals at the current time. We introduced an adaptive layer into the full connection layer of a deep convolution neural network. The objective function of the adaptive layer was designed to minimize the Local Maximum Mean Discrepancy (LMMD) and the prediction error while minimizing the d istance w ithin each c lass (DWC) and maximizing the d istance b etween c lasses w ithin each d omain (DBCWD). We verified the effectiveness of the proposed algorithm on two public datasets. The classification accuracy of the proposed algorithm was higher than other comparison algorithms, and the paired t-test results also showed that the performance of the proposed algorithm was significantly different from that of other algorithms. The results of the confusion matrix and feature visualization showed the effectiveness of the proposed algorithm. Experimental results showed that the proposed algorithm can achieve higher classification accuracy than other algorithms when there was only a small amount of labeled MI EEG data at the current time. It can be promising to be applied to the field of BCI. [ABSTRACT FROM AUTHOR]

Feature extraction networks play a crucial role in classification algorithms. However, most convolutional neural networks use the form of residual block stacking to extract downstream features of images, and this simple form is not sufficient to obtain better classification performance. Therefore, this paper proposes a YOLOX detection framework based on an improved ConvNeXt which incorporates many advantages of Swin-Transformer with efficient classification performance. To address the problems of unstable IOU convergence and inaccurate regression coordinates of the algorithm, this paper adopts Manhattan distance as the new penalty term, and the new loss function combines the advantages of Euclidean distance and Manhattan distance, which has no gradient explosion in the early training period and can drop smoothly to a stable point in the late training period. The proposed algorithm is robust to negative effects such as dust, light, shadows, and occlusions, and is well-suited for masked-face detection in industrial and public scenarios. The proposed model is tested on the VOC dataset and the AIZOO open-source masked-face dataset, obtaining 5.58% AP improvement on the VOC dataset and reaching 96.665% AP on the masked-face dataset, which is better than other algorithms. • Designing a deep neural network to use for masked-face detection. • Anchor-free model is used to avoid time-consuming post-processing. • The improved ConvNeXt network improves classification performance. • Adopting Manhattan distance-based localization loss makes training stable. • The Mosaic data augmentation strategy alleviates the overfitting problem. [ABSTRACT FROM AUTHOR]

Developing 3D Convolutional Neural Networks (CNNs) for medical image segmentation is challenging due to the limited number of available labeled medical images and computational resources. Effective design of a 3D CNN requires a better structure and appropriate hyperparameters. Manually designing a neural network and choosing appropriate hyperparameters requires extensive knowledge, a time-consuming and challenging process, resulting in complex and over-parameterized networks. Neuroevolution is an automatic and effective approach to finding an optimal network structure and hyperparameters. This paper proposes an evolutionary based method using Chameleon Swarm Algorithm (CSA) to develop 3D CNN networks for volume segmentation. We also design a search space to build optimal 3D block-based encoder–decoder structure networks for medical image segmentation. We evaluated the proposed method on two different datasets (CT Spleen and MRI Heart) and compared the results of the obtained model with state-of-the-art models. Furthermore, the comparative results show that our proposed method yields 3D CNN networks with superior performance and fewer parameters than other state-of-the-art models utilized in 3D medical image segmentation. [Display omitted] • Proposed a dynamic 3D CNN for medical image segmentation with fewer parameters. • Implemented search space with various blocks, parameters for encoder–decoder network. • Employed CSA to search the optimal network structure and training hyperparameters. • Results demonstrate the effectiveness of proposed method for 3D segmentation tasks. [ABSTRACT FROM AUTHOR]

Computer-generated holography (CGH) is a technique that aims to produce specific illumination patterns using coherent light. However, traditional CGH algorithms often struggle to balance computational speed with the accuracy of the generated hologram. To address this issue, a non-iterative algorithm named "DL-GSA" is proposed in this paper. DL-GSA combines unsupervised learning in machine learning with convolutional neural networks (CNN) to generate holograms with high accuracy and fixed computational complexity. Simulation experiments reveal that DL-GSA generates hologram patterns faster than the Modified Gerchberg-Saxton algorithm (MGSA) and double-phase retrieval algorithm (DPRA). Furthermore, the average accuracy of the generated holograms is higher than 95 %. These findings suggest that DL-GSA has the potential to significantly enhance the real-time hologram generation capabilities, making it a promising technique for future applications. [ABSTRACT FROM AUTHOR]

Image demosaicking and denoising are the first two key steps of the color image production pipeline. The classical processing sequence has for a long time consisted of applying denoising first, and then demosaicking. Applying the operations in this order leads to oversmoothing and checkerboard effects. Yet, it was difficult to change this order, because once the image is demosaicked, the statistical properties of the noise are dramatically changed and hard to handle by traditional denoising models. In this paper, we address this problem by a hybrid machine learning method. We invert the traditional color filter array (CFA) processing pipeline by first demosaicking and then denoising. Our demosaicking algorithm, trained on noiseless images, combines a traditional method and a residual convolutional neural network (CNN). This first stage retains all known information, which is the key point to obtain faithful final results. The noisy demosaicked image is then passed through a second CNN restoring a noiseless full-color image. This pipeline order completely avoids checkerboard effects and restores fine image detail. Although CNNs can be trained to solve jointly demosaicking–denoising end-to-end, we find that this two-stage training performs better and is less prone to failure. It is shown experimentally to improve on the state of the art, both quantitatively and in terms of visual quality. [ABSTRACT FROM AUTHOR]

This paper proposes CNN based "text in image" steganography using slice encryption algorithm and LWT technique. Generally, cryptography and steganography techniques are used to achieve the covert communication. However, these techniques may not provide good security to the secret data individually. This paper brings the solution by the combination of cryptography, steganography and eight layer CNN. The slice encryption algorithm generates a less length effective ciphertext when compared to the actual length of the readable secret data. This is because; the algorithm uses unique data conversion (DC) and "gray-code" techniques. Next, the new architecture of eight layer CNN is used to generate the feature map in such a way that it accepts the high capacity of ciphertext. Finally, two-level LWT provides the sufficient number of sub-bands which helps to generate high perceptual quality of stego image and good quality metric values. The experimental results shown that the stego image generated from proposed framework have good quality metric values for various payload capacities when compared to existing methods. Finally, the stego image generated with these combinations (Slice, CNN & LWT) yields more payload capacity, high perceptual quality, good quality metric values and can able to withstand for various attacks. [ABSTRACT FROM AUTHOR]

• A small sample pavement crack detection framework based on GAN and CNN. • The image generated by the proposed GAN model can retain the high-frequency information and be directly used in CNN model. • The detection method based on CAM enables the classification model to provide semantic information and expands the detection ability of the classification model. Pavement crack detection methods based on deep learning and computer vision can greatly improve detection efficiency and accuracy, but in many cases the data in training set is lacking or uneven, making it insufficient to train an accurate detection model. This paper proposes a detection method under small samples, which is composed of two steps. First, with a generative adversarial network (GAN) constructed, the small sample data set of pavement cracks taken by unmanned aerial vehicle (UAV) is used as the training set and the GAN model is trained. The best trained model is used for generation of new images. Second, original small-sample data set is expanded by images generated by the GAN model, and a convolutional neural network (CNN) model is constructed at the same time. Then, data set before and after the expansion is trained and tested by the method of transfer learning to verify the effectiveness of expanded data separately. It has been proved that, compared with the unexpanded data set, CNN model trained after expansion improves the test set detection accuracy from 80.75% to 91.61%, which is regarded as a significant improvement. In addition, this paper also uses class activation map (CAM) to visually evaluate CNN model, and expands the detection ability of classification model. [ABSTRACT FROM AUTHOR]

Single-cell mechanics have gained much attention due to its importance in a broad range of biological applications. Different experimental approaches have been used for measuring the mechanical properties of individual cells. However, the technical demands and time-consuming nature of these procedures have limited the throughput of single-cell measurement, necessitating the development of alternative computational approaches. Recently, single-cell deformability can be predicted using a convolutional neural network (CNN) model, shedding a light on using machine learning (ML) algorithms for high-throughput characterizations of single-cell mechanical properties. In this work, we developed a novel ML-based computational framework that can reproduce a physical microfluidic system to investigate the individual cell's deformability. The datasets for the training and testing of our model were generated using high-fidelity fluid-structure interaction (FSI) simulations. Our FSI-based ML approach of adopting CNN algorithms demonstrated a highly accurate prediction for the membrane stiffness of a microcapsule (maximum R2 = 0.98) based on its deformed shape. In this paper, we show that by applying physical constraints including the microcapsule's total surface area and total volume, we were able to build a physics-constrained ML model that possesses better convergence and higher stability during both training and validation. Finally, we found that ML models that used the three-dimensional geometry of the capsule as input could outperform the typical CNN models that relied solely on the two-dimensional images. We expect that this physics-constrained computational framework will serve as a basis for developing future tools for real-time biological applications through the integration of high-fidelity simulations with ML algorithms. • Integration of machine learning (ML) algorithms and fluid dynamics simulations to study individual cell's deformability. • This computational framework predicts accurately microcapsule's membrane stiffness based on its deformed shape. • Applying physical constraints on the ML models achieves better convergences during both training and validation. • CNN models using three-dimensional geometry as input outperform typical CNN models that solely use two-dimensional images. [ABSTRACT FROM AUTHOR]

• A directed convolutional neural network model is proposed • Partial dilation-Global erosion generates resistance against information loss • Several recent approaches for the segmentation of graphemes/characters are discussed • The over-traces issue is addressed for the first time in the literature • MMOs can provide real solutions if a convenient strategy is adopted Due to the nature of Arabic handwriting, segmenting words into characters/graphemes is the most difficult and critical task of the recognition system. The present paper proposes an approach to segment handwritten Arabic words into graphemes based on a directed Convolutional Neural Network (CNN) and Mathematical Morphology Operations (MMO). Arabic script is cursive, which means that almost all graphemes are connected via horizontal links; therefore, a technique to remove links will facilitate the segmentation of graphemes. In general, an MMO such as erosion seems suitable for getting the job done, but since Arabic handwriting is difficult, MMOs cause information loss and suffer from many issues such as diacritics and over-traces, which lead to over/under/bad segmentations. To overcome limitations, the present paper addresses these issues in the following order: the over-traces issue is addressed for the first time in the literature; a robust algorithm for diacritics extraction is provided; and finally, the main segmentation algorithm adopts a strategy based on a Partial Dilation (PD)-Global Erosion (GE) technique to combat the information loss issue. The PD phase amplifies important regions, while GE eliminates links between graphemes. The complementarity between PD and GE facilitates the extraction of graphemes and creates resistance against information loss. To properly tackle these difficult problems, this article exploits the robustness of CNNs, so a new directed CNN model is suggested. The idea is to draw the model's attention to certain targeted features, which are selected according to the nature of the problem addressed. The proposed directed CNN is used in all phases of the segmentation process. The experimental results are very encouraging and show that the proposed directed CNN model outperformed basic CNN in many experiments. The results also reveal that the followed strategy improved the ability of MMOs to perform segmentation and to compete with other approaches in this research area. [ABSTRACT FROM AUTHOR]

• The algorithms of locomotion intention identification for lower limb exoskeleton in recent years are compared and analyzed. • Discussing the relationship between sensors/devices location and type in locomotion intention recognition of lower limb exoskeleton. • Summarizing the connection between the types of sensors and different subjects' tasks. • Analyzing the performance of different algorithms in locomotion intention identification. In recent years, lower limb exoskeletons (LLEs) have received much attention due to the potential to help people with paraplegia regain the ability of upright-legged locomotion. However, one major hindrance to converting prototypes into actual products is the lack of a balance recovery function. Locomotion intentions can be the first step for balance assistance. Therefore, its significance continues to grow. Many researchers focus on this topic, but there is a lack of a general discussion on the research phenomenon. Therefore, the purpose of this work is to systematize these data and benefit future research. This review is divided into two parts, the location of sensors/devices and the evaluation criteria of algorithms, which are the main components of locomotion intentions. We found that sensor/device placement is still concentrated in the lower limbs, but most researchers have found the importance of the chest. The peak power of the signal collected from the chest may be overestimated because it undergoes higher vertical velocity and acceleration during a rotation. However, despite the differences in peak power between the upper and lower back, high correlations were found for the tasks, especially from sitting to standing. Since peak power is based on vertical acceleration and velocity, it can be considered a metric that is more robust to changes in sensor location. Therefore, data acquisition from the chest is effective. In this paper, it is pointed out that sensors placed on the chest may have a tendency to change, as some researchers have realized in the field of locomotion intention recognition. In the evaluation criteria, we also found that deep learning algorithm (such as Back Propagation Artificial Neural Network (BPANN)) is outstanding, and Support Vector Machine (SVM) is the most cost-effective algorithm. In terms of accuracy, sensitivity, and specificity, BPANN achieved nearly 100%. SVM has different types; the best one achieves 98% accuracy, 100% sensitivity, and 100% specificity. But it also has 87.8% accuracy, which is not stable. Convolutional Neural Networks (CNN) can be used for image classification and have an accuracy of around 87%. Compared to the above two algorithms, CNN may have lower performance. Other algorithms also have higher accuracy, sensitivity, and specificity. These evaluation criteria, however, were not all ideal at the same time. Based on these results, we also point out the existing problems. In general, the application of these algorithms to LLE can contribute to its intention recognition, which can be helpful in balancing research. Finally, this can help make LLE more suitable for daily use. [ABSTRACT FROM AUTHOR]

Multi-point geostatistics has been widely used, but it involves many parameters when applying the method. These modeling parameters are usually sensitive to the quality of the model, such as the template size of the simpat algorithm, the number of nodes in the neighborhood searched of the Snesim algorithm, and the multigrid in common use. What kind of evaluation criteria for selecting the value and ranges of the parameters has become an urgent problem to be solved. Some methods to optimize these modeling parameters have been proposed. These methods either statistics the pattern feature distribution of training image, or use some certain spatial structure as the index to quantitatively evaluate the quality of training image reconstruction of stochastic realizations, or use cross-validation to compare the sample matching degree between the realization and the training image. However, these methods have their limitations in optimizing modeling parameters. In this paper, two methods to optimize the modeling parameters of multi-point geostatistics were proposed based on gray level co-occurrence matrix and convolutional neural network. Both methods can optimize the appropriate parameters by analyzing the influence of the parameters on modeling quality. Taking the frequently used parameters of Snesim and simpat as an example, the parameter optimization results of the GLCM-based method and the Deep Learning-based method are very close, which indicates that both methods can effectively extract and distinguish the spatial features of the model. In addition, the relationship between modeling parameters and evaluation indexes accords with the visual features of stochastic models. Therefore, the method proposed in this paper will be helpful to solve the difficulty of parameter setting in multi-point geostatistical modeling. • Two methods to optimize the modeling parameters of multi-point geostatistics are proposed. • The template size in pattern-based methods is optimized. • The search Template Nodes Number in Snesim is optimized. • The multiple grids number is optimized. [ABSTRACT FROM AUTHOR]

• Convolutional Neural Networks can be improved in terms of the classification performance and robustness by using variable weight structures. • Analysis of different data processing methods, models' robustness and statistical properties. • Comparative analysis of variable weight convolutional neural networks and other widely used machine learning techniques. • Medical applications to the classification of seizure phases and types. Deep learning techniques have recently achieved impressive results and raised expectations in the domains of medical diagnosis and physiological signal processing. The widely adopted methods include convolutional neural networks (CNNs) and recurrent neural networks (RNNs). However, the existing models possess static connection weights between layers, which might limit the generalization capability and the classification performance of the models as the weights of different layers are fixed after training. Furthermore, to deal with a large amount of data, a neural network with a sufficiently large size is required. This paper proposes the variable weight convolutional neural networks (VWCNNs), which are a type of network structure employing dynamic weights instead of static weights in their convolutional layers and fully-connected layers. VWCNNs are able to adapt to different characteristics of input data and can be viewed as an infinite number of traditional, fixed-weight CNNs. We will show that the proposed VWCNN structure outperforms the conventional CNN in terms of the classification accuracy, generalization capability, and robustness when the inputs are contaminated by noise. In this paper, VWCNNs are applied to the classification of three seizure phases (seizure-free, pre-seizure and seizure) based on measured electroencephalography (EEG) data. VWCNNs achieve 100% test accuracy and show strong robustness in the classification of the three seizure phases, and thus show the potential to be a useful classification tool for medical diagnosis. Furthermore, the classification of seven types of seizures is investigated in this paper using the world's largest open source database of seizure recordings, TUH EEG seizure corpus. Comparisons with conventional CNNs, RNN, MobileNet, ResNet, DenseNet and traditional machine learning methods including random forest, decision tree, support vector machine, K-nearest neighbours, standard neural networks, and Naïve Bayes are being conducted using realistic test data sets. The results demonstrate that VWCNNs have advantages over other classifiers in terms of classification accuracy and robustness. [ABSTRACT FROM AUTHOR]

Detection transformer (DETR) combines convolutional neural network (CNN) with transformer, providing a more advanced idea. In this paper, an object detection model based on DETR is proposed for fire and smoke detection. Compared with other methods that based on deep learning, the proposed one simplifies the pipeline of detection and builds an end-to-end detector. At the same time, the original DETR model usually requires long training time and large amount of computation, resulting in relatively poor performance in detection speed and accuracy, and it shows to be not friendly for small or early fire detection. Therefore, when designing the proposed network model, a normalization-based attention module is added in the feature extraction stage to highlight the effective features, being beneficial to the process of encoding. A multiscale deformable attention is also used in the encoder–decoder structure, accelerating therefore, the process of convergence during the training of the model, which includes the enhancement of the detection of small objects. Also, considering the cost of computation, the number of layers in the encoder–decoder structure is redefined to reduce the complexity of the model, and that also reduces the requirements of the application equipment. Detailed experiments are conducted on three self-built datasets and two public video sets. The results show that, the proposed method has an excellent performance on all of the datasets considered here. [ABSTRACT FROM AUTHOR]

• Dynamic flexible job shop scheduling problem with machine breakdown is studied. • A two-stage algorithm based on convolutional neural network is proposed. • The improved imperialist competition algorithm is proposed to generate schedules. • A predictive model is proposed to predict the robustness of scheduling. In the actual manufacturing process, the environment of the job shop is complex. There will be many kinds of uncertainties such as random job arrivals, machine breakdowns, order cancellations and other dynamic events. In this paper, an effective two-stage algorithm based on convolutional neural network is proposed to solve the flexible job shop scheduling problem (FJSP) with machine breakdown. A bi-objective dynamic flexible job shop scheduling problem (DFJSP) model with the objective of maximum completion time and robustness is established. In the two-stage algorithm, the first stage is to train the prediction model by convolutional neural network (CNN). The second stage is to predict the robustness of scheduling through the model trained in the first stage. First, an improved imperialist competition algorithm (ICA) is proposed to generate training data. Then, a predictive model constructed by CNN was proposed, and an alternative metric called RMn was developed to evaluate robustness. RMn evaluates that the float time has an effect on the robustness through the information of machine breakdown, workload and float time of the operation. The experimental results show that the proposed two-stage algorithm is effective for solving DFJSP, and RMn can evaluate the robustness of scheduling more quickly, efficiently and accurately. [ABSTRACT FROM AUTHOR]

Although the production process of liquid crystal display model has been automated, the quality detection still depends on manual work. Mura defect is one of the common defects appearing in TFT-LCD modules. Since mura defect is not significantly different from the common background, it is difficult to detect. This paper presents a deep channel attention-based classification network (DCANet), which acts as a powerful feature extractor for object detectors, and proposes an antagonistic training algorithm based on convolution neural network. By the proposed training approach, the deep learning-based object detectors can achieve high accuracy even with a small number of training samples of mura defect. The experimental results show that compared to vanilla training method, the deep learning-based detectors trained by our proposed method could significantly improve their performance on mura defect detection with a few training samples. Even trained on only 600 samples, the mistake rate and miss rate are only 8.08% and 0.267% respectively, which can completely fulfill the enterprise's requirements of 10% and 0.3%. • A lightweight backbone network based upon dense blocks with channel attention (DCANet) is proposed and analyzed. • A data augment method of generate antagonistic sample is presented. • An antagonistic training framework is proposed to train deep learning-based object detectors. • Detailed study for the proposed backbone network and training algorithm. [ABSTRACT FROM AUTHOR]

Autonomous vehicles are the latest engineering research sensations. The development of these needs the collaboration of studies in the automotive and artificial intelligence domains. Self-driving cars face a slew of security issues when left on their own without human supervision. However, the massive volumes of data required for vehicle operation can't be processed rationally by humans. This is where modern intelligence systems can help. They can facilitate clear decision-making regarding the state of the systems. A unique technique based on Machine Learning (ML) procedures is proposed in this paper. Two cluster models have been constructed and coupled with three Convolutional Neural Networks (CNN) models based on Demand-Side Management (DSM). These models are integrated to determine the optimal operating time, stability, and management model. Data from the previous three years have been used in this study. The results show that the proposed CNN and cluster 2 models are the best for energy optimization. The best thing is that this real-time model can meet the demands of large traffic. [Display omitted] [ABSTRACT FROM AUTHOR]

• Analyze the structure of conventional atomic decomposition algorithm (CAD). • Improve the computational efficiency and the accuracy of sub-dictionaries. • Propose an improved CNN-based atomic decomposition algorithm. The atomic decomposition (AD) algorithm for Power Quality Disturbance (PQD) signals can obtain sparser and physically clearer results than the conventional fixed basis decomposition method. However, the conventional atomic decomposition (CAD) algorithm for PQD signals suffers from excessive computational resource consumption and low accuracy of sub-dictionaries selection. In this paper, the CAD algorithm for PQD signals is optimized by introducing a Convolutional Neural Networks based (CNN) sub-dictionary predictor inspired by the PQD signal classification technique. In the optimized algorithm, by using a sub-dictionary predictor to select sub-dictionaries directly, the range of each atomic search is significantly reduced. In addition, the undesirable effects of the intelligent optimization algorithm are mitigated, which match the non-global optimal results in search of sub-dictionaries. Besides, the goal of reducing computation resources is achieved, and the accuracy of sub-dictionaries selection is improved. Finally, the CAD algorithm and optimized CNN-based atomic decomposition (CNN-AD) algorithm are compared and analyzed on the synthetic dataset and the measured dataset from IEEE 1159.2 Working Group on power quality disturbances. Consequently, it is verified that the CNN-AD can reduce the requirement of computation resources and improve the accuracy of sub-dictionaries selection in the decomposition for PQD signals. [ABSTRACT FROM AUTHOR]

Synthetic aperture radar (SAR) is a powerful tool for monitoring sea states in terms of the significant wave height (SWH). Regarding the specific wave mode, to date, the previous empirical models for estimating SWH from SAR data rely on single polarization. In the emerging deep learning era, few published quad-polarized SAR SWH retrieval algorithms have been based on machine learning technique, and whether quad-polarimetry improves the skill of wave height estimation remains a question. Here we propose a deep residual convolutional neural network-based SAR SWH retrieval algorithm in quad-polarization. By collocating WaveWatch III sea state hindcasts and all available archives of quad-polarized Chinese Gaofen-3 SAR imagettes in wave mode, a database with approximately 30,000 matchups was employed to establish our deeply-learned network. The GaoFen-3 significant wave height retrievals were validated against the hindcast dataset independent of training along with altimeter observations. The result of good consistency in terms of a root mean square error of 0.32 m (under sea state conditions of approximately 0.5–7.0 m) outperforms the existing Gaofen-3 wave height retrieval algorithms. Additionally, this paper introduces a discussion about the contribution of polarizations by comparing SWH derived from single-, dual- and quad-polarized deep convolutional neural networks. Single-polarized Gaofen-3 SAR data are found to be sufficient to provide accurate estimates compared to quad-polarization via a deep learning model under moderate sea conditions. Exploitation of SAR quad-polarimetry information will improve SAR wave height retrievals under high sea conditions. • We propose residual networks for estimating wave heights from Gaofen-3. • Our approach outperformed the existing Gaofen-3 wave height retrieval methods. • Single polarization is sufficient for accurately estimating moderate wave heights. • Quad-polarized data enhance the SAR wave estimating capabilities for high states. [ABSTRACT FROM AUTHOR]

• The novel semantic segmentation based visual servoing algorithm is proposed. • The new 4 step algorithm for nonholonomic mobile robot control is developed. • Fully convolutional neural networks are utilized for mobile robot visual servoing. • The approach is validated by using both real mobile robot RAICO and 3D simulation. In the interest of developing an intelligent manufacturing environment with an agile, efficient, and optimally utilized transportation system, mobile robots need to achieve a certain level of autonomy as they play an important role in carrying out transportation tasks. Bearing this in mind, in the paper we propose a novel stereo visual servoing method for nonholonomic mobile robot control based on semantic segmentation. Semantic segmentation provides a rich body of information required for an adequate decision-making process in a clustered, dynamic, and ever-changing manufacturing environment. The innovative idea behind the new visual servoing system is to utilize semantic information of the scene for visual servoing, as well as for other mobile robot tasks, such as obstacle avoidance, scene understanding, and simultaneous localization and mapping. Semantic segmentation is carried out by exploiting fully convolutional neural networks. The new visual servoing algorithm utilizes an intensity-based image registration procedure, which results in the image transformation matrix. The transformation matrix encompasses the relations of images taken at the current and desired pose, and that information is directly used for visual servoing. The developed algorithm is deployed on our own developed wheeled differential drive mobile robot RAICO (Robot with Artificial Intelligence based COgnition). The experimental evaluation is carried out in the 3D simulation environment and in the laboratory model of the real manufacturing environment. The experimental results show that the accuracy of the proposed approach is improved when compared to the state-of-the-art approaches while being robust to the partial occlusions of the scene and illumination changes. [ABSTRACT FROM AUTHOR]

Face sketch synthesis plays a crucial role in face recognition for law enforcement applications. However, the current face sketch synthesis approaches generate sketches from photos based on a model trained by a certain database that is usually collected from individuals of the same ethnicity, and therefore such sketches merely inherit distinct facial distributions (shape and texture) of this database. This also makes such models inapplicable for real-world applications which mainly include multiple photo variations such as pose, lighting, skin color, and ethnic origin. In this paper, a unified face sketch synthesis model considering ethnicity issue as well as photo variations is proposed. A new deep learning scheme is designed to handle the generic visual representation and global structure of the face. Towards the final objective, the recent success of deep residual blocks is exploited and incorporated into a plain feedforward network, termed as DResNet, to learn a regression model for face sketch synthesis. A heterogeneous database containing photos with lighting, ethnicity, hair and skin variations is utilized for training the DResNet model. Extensive subjective and objective evaluations showed the superiority of the proposed DResNet method on state-of-the-art face sketch synthesis methods. Experimental results also demonstrated that the proposed DResNet method can be generalized to face sketch synthesis for real-world applications. • A face sketch synthesis algorithm on the basis of Deep Residual Network is proposed. • A plain feedforward convolutional network is introduced to enhance the feature maps. • Generic visual representation and global structure of the face are clearly presented. • A heterogeneous database is used to evaluate the proposed algorithm. • The proposed algorithm is performed favorably with real-world face photos. [ABSTRACT FROM AUTHOR]

Male infertility is a disease that affects approximately 7% of men. Sperm morphology analysis (SMA) is one of the main diagnosis methods for this problem. However, manual SMA is an inexact, subjective, non-reproducible, and hard to teach process. Therefore, in this paper, we introduce a novel automatic SMA technique that is based on the neural architecture search algorithm, named Genetic Neural Architecture Search (GeNAS). For this purpose, we used a collection of images termed MHSMA dataset, which contains 1540 sperm images that have been collected from 235 patients with infertility problems. In detail, GeNAS consists of a special genetic algorithm that acts as a meta-controller which explores the constrained search space of plain convolutional neural network architectures. Every individual of this genetic algorithm is a convolutional neural network trained to predict morphological deformities in different segments of human sperm (head, vacuole, and acrosome). The fitness of each individual is calculated by a novel proposed method, named GeNAS Weighting Factor (GeNAS-WF). This technique is specially designed to evaluate the fitness of neural networks which, during their learning process, validation accuracy highly fluctuates. To speed up the algorithm, a hashing method is practiced to save each trained neural architecture fitness, so we could reuse them during fitness evaluation. In terms of running time and computational power, our proposed architecture search method is far more efficient than most of the other existing neural architecture search algorithms. Moreover, whereas most of the existing neural architecture search algorithms are designed to work well with well-prepared benchmark datasets, the overall paradigm of GeNAS is specially designed to address the challenges of real-world datasets, particularly shortage of data and class imbalance. In our experiments, the best neural architecture found by GeNAS has reached an accuracy of 91.66%, 77.33%, and 77.66% in the vacuole, head, and acrosome abnormality detection, respectively. In comparison to other proposed algorithms for MHSMA dataset, GeNAS achieved state-of-the-art results. • A novel neural architecture search based on Genetic Algorithm. • A novel neural architecture encoding for exploration of search space. • A specific Crossover and Mutation operations for neural architecture search. • Work on imbalanced datasets, low resolution images, and non-stained sperms. • State-of-the-art accuracy & Real-time sperm morphological analysis. [ABSTRACT FROM AUTHOR]

• Unified approach to recognize egocentric hand gesture and detect fingertips. • Binary output of learning network as visible fingers in a gesture. • Introducing an ensemble of direct regression to predict fingertip position. • Experiments on known database and real-life images along with VR application. • Achieved highly accurate gesture classification and fingertip position. Head-mounted device-based human-computer interaction often requires egocentric recognition of hand gestures and fingertips detection. In this paper, a unified approach of egocentric hand gesture recognition and fingertip detection is introduced. The proposed algorithm uses a single convolutional neural network to predict the probabilities of finger class and positions of fingertips in one forward propagation. Instead of directly regressing the positions of fingertips from the fully connected layer, the ensemble of the position of fingertips is regressed from the fully convolutional network. Subsequently, the ensemble average is taken to regress the final position of fingertips. Since the whole pipeline uses a single network, it is significantly fast in computation. Experimental results show that the proposed method outperforms the existing fingertip detection approaches including the Direct Regression and the Heatmap-based framework. The effectiveness of the proposed method is also shown in-the-wild scenario as well as in a use-case of virtual reality. [ABSTRACT FROM AUTHOR]

Most of the available recommender systems focus on the accuracy of recommendations. As a result, their recommendations are often popular and very close to user preferences, which make them repetitious and predictable, hence adversely affecting user satisfaction. Recent studies on recommender systems, however, aim for factors beyond accuracy as accuracy alone cannot ensure the satisfaction of all users. One of the most important criteria beyond the accuracy is serendipity, which includes relevant, unexpected, and novel recommendations that cannot be easily discovered by users themselves. In this paper, a Convolutional Neural Network (CNN) is integrated with the Particle Swarm Optimization (PSO) algorithm to generate serendipitous recommendations. The proposed method is based on the focus shift points, consisting of unexpectedness and relevance parameters. In this approach, these points are considered as the factors showing whether recommendations are serendipitous. The CNN is employed to predict the focus shift points for each user. Then, the PSO is utilized to search for recommendations close to the predicted focus shift points and generate the list of candidate recommendations. After that, the Serendipitous Personalized Ranking (SPR) method is employed to re-rank the candidate recommendations and generate the final list. According to the evaluation results, the proposed approach outperforms other state-of-the-art methods in SRDP, Hit Ratio, and NDCG factors. [ABSTRACT FROM AUTHOR]

Nowadays, traffic situation prediction is a major concern. Travel time index prediction is one of the representative tasks in traffic situation prediction. Most research of the current prediction focuses on road-level speed, but the travel time index prediction of complex road network at the regional level has not been well solved. In this paper, we proposed a novel model with convolutional neural network called Seq2Img-DRNET. The sequence data with 10 min interval are converted into image data as the input of our model. The model is constructed by fusing dense connection network and residual network to solve the travel time index prediction of complex road network at the regional level. Taking two actual traffic networks of commercial district and scenic district in Chengdu as the example, we compared our model with common sequential model and convolution neural network model, and the results show that our model can effectively capture the spatial–temporal relationship characteristics of complex road network at the regional level and make accurate prediction. • Our method transforms time series data into image data for spatiotemporal information fusion. • A convolutional neural network model is constructed to accurately predict the travel time index. • Aiming at urban business districts and scenic spots prone to traffic congestion. • The prediction accuracy is significantly higher than the existing baseline model. [ABSTRACT FROM AUTHOR]

[Display omitted] • A contribution to drug repurposing using deep learning and optimization is proposed. • A novel deep learning model is developed to predict Drug-Target binding affinities. • The proposed model integrates a Convolutional Neural Network on the top of a BiLSTM. • The Hyperparameters of the model are set using a Differential Evolution algorithm. • The Differential Evolution-based learning framework outperformed baseline methods. The identification of drug-target interactions (DTIs) is an important process in drug repositioning and drug discovery. However, it is very expensive and time-consuming to determine all possible DTIs with experimental approaches. Most existing machine learning-based methods formulate the DTIs prediction problem as a binary classification problem. Nevertheless, the lack of experimentally validated negative samples results in imbalanced class distribution within the datasets, which may have a negative influence on the DTI prediction performance. Casting DTI prediction task as a regression problem seems an interesting alternative to avoid this issue especially with the recent increase in protein structural data and DTI datasets. Within this context, a twofold contribution is described in this paper. First, we propose a novel deep learning model for predicting drug-target binding affinities called "Convolution Neural Network with Attention-based bidirectional Long Short-Term Memory network" (CNN-AbiLSTM), which combines a CNN with an attention-based biLSTM. Second, building a powerful hybrid CNN-AbiLSTM model can be highly complicated and requires a suitable setting of the model's hyperparameters. To handle this problem, we propose an evolutionary algorithm-based framework more specifically a Differential Evolution (DE) algorithm to find the optimal configuration of the proposed model. Experimental results show that the proposed DE-based CNN-AbiLSTM model achieves better performance compared with baseline methods. [ABSTRACT FROM AUTHOR]

• A multi-task fully deep convolutional neural network is proposed for contactless fingerprint minutiae extraction. • A new loss function is proposed to jointly learn the tasks of minutiae detection and its direction regression from the whole fingerprints. • The proposed method operates directly on the gray scale contactless fingerprints without any preprocessing. • Experiments on three datasets have shown the effectiveness of the proposed algorithm. With the outbreak and wide spread of novel coronavirus (COVID-19), contactless fingerprint recognition has attracted more attention for personal recognition because it can provide significantly higher user convenience and hygiene than the traditional contact-based fingerprint recognition. However, it is still challenging to achieve a highly accurate recognition due to the low ridge-valley contrast and pose variances of contactless fingerprints. Minutiae points are a kind of ridge flow discontinuities, and robust and accurate extraction is an important step for most automatic fingerprint recognition algorithms. Most of existing methods are based on two stages which locate the minutiae points first and then compute their directions. The two-stage method cannot make full use of location and direction information. In this paper, we propose a multi-task fully deep convolutional neural network for jointly learning the minutiae location detection and its corresponding direction computation which operates directly on the whole gray scale contactless fingerprints. The proposed method consists of offline training and online testing stages. In the training stage, a fully deep convolutional neural network is built for the tasks of minutiae detection and its direction regression, with an attention mechanism to make the direction regression branch concentrate on the minutiae points. A new loss function is proposed to jointly learn the tasks of minutiae detection and its direction regression from the whole fingerprints. In the testing stage, the trained network is applied on the whole contactless fingerprint to generate the minutiae location and direction maps. The proposed multi-task leaning method performs better than the individual single task and it operates directly on the raw gray-scale contactless fingerprints without preprocessing. The results on three contactless fingerprint datasets show the proposed algorithm performs better than other minutiae extraction algorithms and the commercial software. [ABSTRACT FROM AUTHOR]

Image enhancement can accentuate image feature and is necessary process in image processing. This work focuses on fusing multi-exposure image sequences low-light image enhancement. Inspired by the classical non-local means in computer vision, we proposed an improved deep neural network framework with attentions for image enhancement. Firstly, the original image was preprocessed in different dimensions. we get the edge images using an edge extracted algorithm and fusion multi exposed images to get an better initial images based on fully convolutional neural network with position and channel attention mechanism. Secondly, the head network is constructed by fully convolutional neural network. For capturing long-range dependencies between features maps, we designed a non-local attention module for head network to get better enhancement image. Finally, emerging the original images, edge image and fusion image as the input of the head network, it can enhance the images to get high-quality images. Experiments show that our framework proposed in this paper is effective and the attention mechanism play a significant hole in the network. • Designed channel and space attention mechanism to fusion multi exposure images. • Designed non-local attention mechanism in feature maps domain for enhancement network. • Proposed an emerging deep neural network framework with multiple attention mechanism. [ABSTRACT FROM AUTHOR]

As an essential subproblem in the face recognition field, occluded face recognition has received considerable attention in recent years. However, satisfactory recognition accuracy for occluded faces has not been achieved. Therefore, in order to solve this problem, this paper proposes a convolutional neural network that has multidimensional serial feature extraction modules for occluded faces and uses the deep learning method to improve the recognition rate. In order to improve the expression ability of the network, the method extracts features from two dimensions — space and channel — and built a multidimensional serial feature extraction module. By "multiscale" and "dependence" processing features, the generalization ability was significantly improved. Through the experimental test, the recognition accuracy of the multidimensional feature network (MFNet) reached 90.35%. This accuracy is 17.89% (72.36%) better than that of the traditional algorithm. Meanwhile, compared with other convolutional neural networks, this network has different degrees of improvement, which are 0.68% (ArcFace), 2.14% (Visual Geometry Group Face), and 4.83% (DeepID3). [ABSTRACT FROM AUTHOR]

Automatic liver and tumour segmentation in CT images are crucial in numerous clinical applications, such as postoperative assessment, surgical planning, and pathological diagnosis of hepatic diseases. However, there are still a considerable number of difficulties to overcome due to the fuzzy boundary, irregular shapes, and complex tissues of the liver. In this paper, for liver and tumor segmentation and to overcome the mentioned challenges a simple but powerful strategy is presented based on a cascade convolutional neural network. At the first, the input image is normalized using the Z-Score algorithm. This normalized image provides more information about the boundary of tumor and liver. Also, the Local Direction of Gradient (LDOG) which is a novel encoding algorithm is proposed to demonstrate some key features inside the image. The proposed encoding image is highly effective in recognizing the border of liver, even in the regions close to the touching organs. Then, a cascade CNN structure for extracting both local and semi-global features is used which utilized the original image and two other obtained images as the input data. Rather than using a complex deep CNN model with a lot of hyperparameters, we employ a simple but effective model to decrease the train and testing time. Our technique outperforms the state-of-the-art works in terms of segmentation accuracy and efficiency. [ABSTRACT FROM AUTHOR]

• A DcCNN is used to extract fault features from PVA normalized current and voltage. • A feature selection structure is proposed to evaluate extracted features. • A penalty is developed to train the proposed model in an end-to-end way. • Results show that the DcCNN can extract high-discriminative features. • The DcCNN has superior performance for different faults even LLF under complex conditions. The effective fault diagnosis algorithm for the DC side photovoltaic (PV) array of a PV system (PVS) plays an important role in the operation efficiency and safety for PV power plants. But for fault diagnosis models it may fail to diagnose PV array (PVA) faults without detailed and quite fine fault features, especially line-line faults (LLF) occurring in the PVS that works under complex working conditions like low irradiance conditions and LLF with fault impedance. To address these challenges, this paper proposes a fault diagnosis scheme to diagnose different PVA faults using a proposed Dual-channel Convolutional Neural Network (DcCNN), which is able to automatically extract features and weight these features for fault classification. The important and fine features from the current and voltage electrical time series graph (ETSG) are extracted respectively by DcCNN in a double input way. Then, a proposed feature selection structure (FSS) is designed to improve the proposed fault diagnosis model capacity for diagnosing PVA faults under various conditions, including LLF, partial shading condition (PSC) and open circuit faults (OCF). Comparing to manually designed features, FSS not only helps DcCNN extract important features from PVA current and voltage automatically but also evaluates extracted features for further classification of DcCNN. Moreover, in the training stage, a proposed penalty is applied on DcCNN to constrain FSS, resulting in its sparse weight distribution. A comprehensive experiment based on a laboratory roof grid connected PVS is conducted. The results demonstrate the superior performance of the proposed approach compared with other algorithms as it can extract high-discriminative features from PVA current and voltage for different PVA faults, which is also effective on diagnosing LLF under low irradiance conditions and LLF with fault impedance. [ABSTRACT FROM AUTHOR]

• An end-to-end detection algorithm for layer-wise multi-defect of powder bed is proposed, which can detect defects in a single image with an accuracy of at least 97% within 1.15 s. • The influence of lighting strategy on the detection of powder bed defect is studied, and the optimized lighting strategy is proposed. • An adaptive multi-defect segmentation algorithm for defect extracting is proposed, which can suppress noise and extract defects in a single image in 15.67 ms with a detection rate of 99.44% • The proposed defect detection algorithm can select different convolution neural networks for defect classification according to the task needs and resource constraints, which has good universality and application prospects. Powder bed defects usually inevitably appear in the process of powder spreading during laser powder bed fusion owing to the characteristics of the powder material and the performance of the spreading equipment. This may lead to instability with regard to subsequent processes and the quality of the final part. Defect detection based on the imaging process is an effective way to achieve non-contact, efficient, and accurate online monitoring, and it has received widespread attention. In this paper, an imaging method for online collection of powder bed is proposed based on the experiments of various lighting strategies, whose influences on defect analysis are evaluated. Subsequently, an adaptive segmentation algorithm for defect extraction is proposed that automatically searches for the best threshold by evaluating the gray histogram of the powder bed image. Finally, different convolution neural networks were applied to implement the classification of the defects, and their performances were evaluated and compared. The results of the on-site experiments demonstrate that the proposed method has good accuracy and efficiency in the multi-defect detection of a powder bed. [ABSTRACT FROM AUTHOR]

The drainage system is an important part of civil infrastructure. However, the underground sewage pipe will gradually suffer from defects over time, such as tree roots, deposits, infiltrations and cracks, which heavily affect the performance of sewage pipes. Therefore, it is significant to timely inspect the condition of sewage pipes. Closed-circuit television (CCTV) inspection is a commonly employed underground infrastructure inspection technology requiring engineering experience that can be subjective and inefficient. Nowadays, object detection based on convolutional neural network (CNN) can automatically detect defects, showing high potential for improving inspection efficiency. This paper proposed an improved CNN-based You Only Look Once version 3 (YOLOv3) method for automatic detection of sewage pipe defects, where the improvements are mainly involved in loss function, data augmentation, bounding box prediction and network structure. Experiment results demonstrate that the improved model outperforms Faster R-CNN and YOLOv3, achieving a mean average precision (mAP) value of 92%, which is higher than the existing research on automatic detection of sewage pipe defects. • Proposed an improved YOLO-based method for automatic detection of sewage defects. • The proposed new data augmentation method improved defect detection accuracy. • Applied three different loss functions and compared their detection performance. • Our new architecture outperformed previous research for detecting sewer defects. [ABSTRACT FROM AUTHOR]

• An algorithm for dissected aorta segmentation based on 3-D and 2-D convolutional neural networks. • A 3-D CNN model identifies the proximal and distal regions of the dissected aorta data. • A 2-D CNN model for aorta segmentation which incorporates boundary information. Aortic dissection is a severe cardiovascular pathology in which an injury of the intimal layer of the aorta allows blood flowing into the aortic wall, forcing the wall layers apart. Such situation presents a high mortality rate and requires an in-depth understanding of the 3-D morphology of the dissected aorta to plan the right treatment. An accurate automatic segmentation algorithm is therefore needed. In this paper, we propose a deep-learning-based algorithm to segment dissected aorta on computed tomography angiography (CTA) images. The algorithm consists of two steps. Firstly, a 3-D convolutional neural network (CNN) is applied to divide the 3-D volume into two anatomical portions. Secondly, two 2-D CNNs based on pyramid scene parsing network (PSPnet) segment each specific portion separately. An edge extraction branch was added to the 2-D model to get higher segmentation accuracy on intimal flap area. The experiments conducted and the comparisons made show that the proposed solution performs well with an average dice index over 92%. The combination of 3-D and 2-D models improves the aorta segmentation accuracy compared to 3-D only models and the segmentation robustness compared to 2-D only models. The edge extraction branch improves the DICE index near aorta boundaries from 73.41% to 81.39%. The proposed algorithm has satisfying performance for capturing the aorta structure while avoiding false positives on the intimal flaps. [ABSTRACT FROM AUTHOR]