Search

This paper introduces a novel method to spontaneously predict displacement time histories and hysteresis curves of bridge lead rubber bearings under seismic loads and axial forces. The method leverages a stacked convolutional‐bidirectional Cuda Long Short Term Memory network, enhanced with multi‐head attention, skip connections, exponential learning rate scheduler, and a hybrid activation function to improve performance. The framework utilizes the functional application programming interface provided by the Python Keras library to build a model that takes input features such as horizontal and vertical ground accelerations, actuator loads in both lateral and vertical directions, and the superstructure mass. The effectiveness of the deep learning model is evaluated using a considerable experimental dataset of 53 real‐time hybrid simulations, spanning various earthquake intensities and superstructure masses (Chi‐Chi: 15 scenarios, El Centro: 15 scenarios, Kobe: 13 scenarios, and Northridge: 10 scenarios). Initially, Northridge earthquake data serves as unseen data, while the rest is used for training and validation. In a subsequent trial, the unseen data is centered on Kobe earthquake scenarios. By employing a hybrid loss function merging mean square and mean absolute errors, the model exhibits a substantial correlation of over 83% between predicted displacement time series and empirical measurements for the unseen data. In summary, the proposed model offers miscellaneous benefits, including time and cost savings in experimental efforts by decreasing the need for additional tests. It further delivers a swift and precise insight into the bridge bearing performance and its energy dissipation, facilitating timely and accurate bridge design in different scenarios for engineers. [ABSTRACT FROM AUTHOR]

Accurate segmentation of skin lesion regions is a key step in clinical diagnosis and analysis. Aiming at the poor segmentation effect of the existing networks in skin lesion regions due to the presence of variable size, irregular shape, fuzzy boundaries and obscured lesion regions, a multi-scale and boundary fusion network (MSBF-Net) for skin lesion region segmentation is proposed by improving the original structure based on the U-Net. Specifically, firstly, a split pooling (SplitPool) module is proposed, which effectively solves the problem of spatial information loss while reducing the image resolution. Secondly, a full-scale feature fusion (FSFF) module is proposed, which effectively solves the problem that the previous methods only fuse the deep features to the shallow features, while ignoring the contribution of the detail information in the more shallow features to the network segmentation decision. Meanwhile, the original jump connections of U-Net are redesigned to provide richer contextual information for the decoder. Finally, sub-paths for enhancing the network’s ability to learn boundary features are proposed, and the boundary fusion (BF) module is introduced to fuse the prediction results of the main paths and sub-paths, which effectively solves the problems of irregular shape and boundary ambiguity of the lesion region. Dice and JI reach 90.12% and 83.61% on the ISIC2018 dataset, which are 1.13 percentage points and 1.62 percentage points higher than the baseline network, respectively; Dice and JI reach 94.72% and 90.18% on the PH2 dataset, which are 1.49 percentage points and 2.17 percentage points higher than the baseline network, respectively. Experimental results show that MSBF-Net significantly improves the accuracy of skin lesion region segmentation and exceeds the existing state-of-the-art methods in several indices, further validating the effectiveness of the method.

Copyright of Journal of Frontiers of Computer Science & Technology is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Long short-term memory (LSTM) has proven effective in modeling sequential data. However, it may encounter challenges in accurately capturing long-term temporal dependencies. LSTM plays a central role in speech enhancement by effectively modeling and capturing temporal dependencies in speech signals. This paper introduces a variable-neurons-based LSTM designed for capturing long-term temporal dependencies by reducing neuron representation in layers with no loss of data. A skip connection between nonadjacent layers is added to prevent gradient vanishing. An attention mechanism in these connections highlights important features and spectral components. Our LSTM is inherently causal, making it well-suited for real-time processing without relying on future information. Training involves utilizing combined acoustic feature sets for improved performance, and the models estimate two time–frequency masks—the ideal ratio mask (IRM) and the ideal binary mask (IBM). Comprehensive evaluation using perceptual evaluation of speech quality (PESQ) and short-time objective intelligibility (STOI) showed that the proposed LSTM architecture demonstrates enhanced speech intelligibility and perceptual quality. Composite measures further substantiated performance, considering residual noise distortion (Cbak) and speech distortion (Csig). The proposed model showed a 16.21% improvement in STOI and a 0.69 improvement in PESQ on the TIMIT database. Similarly, with the LibriSpeech database, the STOI and PESQ showed improvements of 16.41% and 0.71 over noisy mixtures. The proposed LSTM architecture outperforms deep neural networks (DNNs) in different stationary and nonstationary background noisy conditions. To train an automatic speech recognition (ASR) system on enhanced speech, the Kaldi toolkit is used for evaluating word error rate (WER). The proposed LSTM at the front-end notably reduced WERs, achieving a notable 15.13% WER across different noisy backgrounds. [ABSTRACT FROM AUTHOR]

Scene text image super-resolution is an interesting and challenging task which aims to enhance the spatial resolution of low-resolution text images in the wild, and consequently improve the image visual quality and boost the performance of real-world text-related applications. However, most of previous super-resolution methods ignore the important specific characteristics of text patterns and regard scene text images as natural scene images. In this paper, a novel deep convolutional-based architecture is specifically proposed for the super-resolution of scene text images. In order to recover fine details of low-resolution characters, the proposed architecture has been carefully designed and its main specificities are three-fold: (1) the introduction of multi-scale features extraction by incorporating parallel convolutional layers in order to preserve both local and global high-frequency components that encapsulate the intricate details of characters' patterns. This strategy allows the proposed method to capture fine nuances in the visual representation of characters, enhancing the richness of extracted features. (2) the integration of skip connections through convolutional layers. This strategic design choice facilitates the seamless flow of information from lower to higher layers of the deep architecture, allowing sequential information about text patterns to be preserved more effectively. (3) the proposition of a specialized network in network-based reconstruction within our architecture to recover high-resolution text details from the collected features. Such a network paradigm minimizes information loss and enhances the proposed method's ability to discern and reconstruct fine textual details. These design elements collectively empower our super-resolution method to excel in analyzing fine text patterns for effective high-resolution reconstruction, providing a comprehensive solution for the challenging task of recovering fine details in low-resolution characters. Quantitative and qualitative evaluations on four well-known benchmarks, including the SVT, IIIT5k, IC03 and ICDAR2015-TextSR datasets, prove the efficiency of our proposal whose performance surpasses those of different state-of-the-art super-resolution methods. [ABSTRACT FROM AUTHOR]

Road extraction is a crucial task that requires high-resolution remote sensing images with wide-ranging applications in urban planning, navigation, and autonomous vehicles. However, this task is challenged by complex road structures and the need to capture long-range dependencies. RoadTransNet is a new road extraction architecture that aims to solve these problems that making the power of the Swin Transformer and Feature Pyramid Network (FPN) while introducing Transformer-like attention mechanisms. RoadTransNet combines a robust convolutional backbone, inspired by the Swin Transformer, with an FPN to capture multi-scale features effectively. The Transformer-like attention mechanisms, including multi-head self-attention and cross-attention, enable the network to represent context information on a local and global scale, ensuring accurate road extraction. The skip connections facilitate gradient flow, preserving fine details, and decoding layers transform extracted features into precise road predictions. Our experiments are conducted using the RoadTransNet, which is subject to rigorous assessment on the following datasets: the DeepGlobe road extraction challenge Dataset and the CHN6-cUG roads dataset. The outcomes indicate its superior performance in achieving high-level metrics of precision and recall, as well as achieving high F1 scores and IoU. The comparative evaluations performed against traditional methods showcase RoadTransNet's ability to capture complex road structures and long-range dependencies. The RoadTransNet stands as a comprehensive solution for the extraction of roads in high-resolution remote sensing images, offering promising opportunities for improving urban planning, navigation systems, and autonomous vehicle technologies. Its success lies in the synergy of convolutional and transformer-based architectures, paving the way for advanced remote sensing applications in smart cities and others. [ABSTRACT FROM AUTHOR]

One of the major threats in computer vison and image processing areas is image tarnishing due to noisy pixels. Image tarnishing occurs when clean images are subjected to multiple feature changes, length transmission, or due to improper image capturing. The growth of Autoencoders in Artificial Neural Networks has paved way to efficient noise removal. In this work, an architecture oriented analysis of stacked denoising autoencoders is implemented by considering two perspectives of noise, namely noise as an inverse problem and noise as a residual problem. Autoencoders have been modelled by considering the approaches: Shallow Model, Deep-Series and Parallel Series models. Autoencoders with Skip Connections and Denoising CNNs have also been implemented and analyzed in order to propose an efficient, optimized and best working model for denoising. The dataset used is Common Object in Context (COCO) and the performance of the model is evaluated based on Structural Similarity Index (SSIM) and Peak Signal to Noise Ratio (PSNR). The analysis has presented the impact of the model structures of the autoencoders, which implicitly tend to treat noise as an inverse problem. [ABSTRACT FROM AUTHOR]

The proliferation of Convolutional Neural Network-based applications in recent years has necessitated the development of effective strategies for information and gradient flow within a network. As networks become deeper, it has been observed that later layers often experience a loss of information from earlier layers, leading to increased difficulties in network training. To address this challenge, this research proposes a progressive research approach to information flow, utilizing a “multi-scale skip connection” method to pass information from early layers to later layers. Specifically, this technique is applied within the context of a DenseNet architecture, where information from early convolutional layers is passed to the pooling layer after each dense block, and information from each dense block is passed to the pooling layer at the end of the subsequent dense block. The effectiveness of this approach is demonstrated through extensive experimental evidence, which indicates that the proposed strategy has the potential to improve accuracy with increased depth, improving feature propagation, and encouraging feature reuse. Furthermore, the proposed strategy is shown to be more accurate and efficient, in terms of both model accuracy and computational resources, on four highly competitive benchmark datasets: CIFAR-10, CIFAR-100, SVHN, and Imagenette. Our model demonstrates superior performance across these datasets, showcasing significant advancements in accuracy and efficiency. On CIFAR-10, with a feature map growth rate of 12 and depth of 64, the proposed network achieves a 4% lower error rate (5.99% vs. 6.24%) and uses 7.6% fewer operations (0.12 vs. 0.13 GMAC) compared to DenseNet. On CIFAR-100, with a growth rate of 12 and depth of 82, it achieves a 5.12% lower error rate (21.33% vs. 22.48%) and uses 6.25% fewer operations (0.15 vs. 0.16 GMAC). Overall, on CIFAR-10, our network achieves an average error rate reduction of 3.34%, and on CIFAR-100, an average reduction of 2.93% compared to DenseNet. Additionally, on Imagenette, our models show an average accuracy improvement of 1.39%. These results are further validated through comparisons with other state-of-the-art networks and strategies.

Regularly detect the corrosion of metal structures and take countermeasures according to the degree of corrosion, which can reduce the potential safety hazards and avoid unnecessary economic losses. This paper presents a deep learning-based metal corrosion detection method, which is used to accurately segment the corrosion regions. This model incorporates Hybrid Attention Residual Block (HARB) and External Skip Connections (ESC) in the basic architecture of U-Net. HARB can reweight the features so that the network pays more attention to the corrosion regions and ignores other irrelevant regions. Shallow features and deep features are fused using ESC to make the features of network learning more effective and comprehensive, which is conducive to improving the segmentation accuracy. In order to further improve the semantic segmentation performance, the Attention Gate Residual Block (AGRB) is proposed, which can optimize segmentation by strengthening the regions to be segmented and reducing the activation value of the background, while overcoming the semantic and scale inconsistencies between input features. In this paper, comprehensive comparative experiments are conducted on the Metal Corrosion dataset. The experimental results show that the F1 and mIoU of the proposed method are 91.21% and 89.56%, respectively. Compared with U-Net, HEU-Net achieves a 7.91% improvement in F1 and a 10.16% improvement in mIoU. These results show that our method outperforms state-of-the-art models and can achieve better semantic segmentation results. [ABSTRACT FROM AUTHOR]

Video anomaly detection is a critical component of intelligent video surveillance systems, extensively deployed and researched in industry and academia. However, existing methods have a strong generalization ability for predicting anomaly samples. They cannot utilize high-level semantic and temporal contextual information in videos, resulting in unstable prediction performance. To alleviate this issue, we propose an encoder–decoder model named SMAMS, based on spatiotemporal masked autoencoder and memory modules. First, we represent and mask some of the video events using spatiotemporal cubes. Then, the unmasked patches are inputted into the spatiotemporal masked autoencoder to extract high-level semantic and spatiotemporal features of the video events. Next, we add multiple memory modules to store unmasked video patches of different feature layers. Finally, skip connections are introduced to compensate for crucial information loss caused by the memory modules. Experimental results show that the proposed method outperforms state-of-the-art methods, achieving AUC scores of 99.9%, 94.8%, and 78.9% on the UCSD Ped2, CUHK Avenue, and Shanghai Tech datasets. [ABSTRACT FROM AUTHOR]

Some recent studies show that filters in convolutional neural networks (CNNs) have low color selectivity in datasets of natural scenes such as Imagenet. CNNs, bio-inspired by the visual cortex, are characterized by their hierarchical learning structure which appears to gradually transform the representation space. Inspired by the direct connection between the LGN and V4, which allows V4 to handle low-level information closer to the trichromatic input in addition to processed information that comes from V2/V3, we propose the addition of a long skip connection (LSC) between the first and last blocks of the feature extraction stage to allow deeper parts of the network to receive information from shallower layers. This type of connection improves classification accuracy by combining simple-visual and complex-abstract features to create more color-selective ones. We have applied this strategy to classic CNN architectures and quantitatively and qualitatively analyzed the improvement in accuracy while focusing on color selectivity. The results show that, in general, skip connections improve accuracy, but LSC improves it even more and enhances the color selectivity of the original CNN architectures. As a side result, we propose a new color representation procedure for organizing and filtering feature maps, making their visualization more manageable for qualitative color selectivity analysis. [ABSTRACT FROM AUTHOR]

Acute lymphoblastic leukemia (ALL) is a life-threatening hematological malignancy that requires early and accurate diagnosis for effective treatment. However, the manual diagnosis of ALL is time-consuming and can delay critical treatment decisions. To address this challenge, researchers have turned to advanced technologies such as deep learning (DL) models. These models leverage the power of artificial intelligence to analyze complex patterns and features in medical images and data, enabling faster and more accurate diagnosis of ALL. However, the existing DL-based ALL diagnosis suffers from various challenges, such as computational complexity, sensitivity to hyperparameters, and difficulties with noisy or low-quality input images. To address these issues, in this paper, we propose a novel Deep Skip Connections-Based Dense Network (DSCNet) tailored for ALL diagnosis using peripheral blood smear images. The DSCNet architecture integrates skip connections, custom image filtering, Kullback–Leibler (KL) divergence loss, and dropout regularization to enhance its performance and generalization abilities. DSCNet leverages skip connections to address the vanishing gradient problem and capture long-range dependencies, while custom image filtering enhances relevant features in the input data. KL divergence loss serves as the optimization objective, enabling accurate predictions. Dropout regularization is employed to prevent overfitting during training, promoting robust feature representations. The experiments conducted on an augmented dataset for ALL highlight the effectiveness of DSCNet. The proposed DSCNet outperforms competing methods, showcasing significant enhancements in accuracy, sensitivity, specificity, F-score, and area under the curve (AUC), achieving increases of 1.25%, 1.32%, 1.12%, 1.24%, and 1.23%, respectively. The proposed approach demonstrates the potential of DSCNet as an effective tool for early and accurate ALL diagnosis, with potential applications in clinical settings to improve patient outcomes and advance leukemia detection research. [ABSTRACT FROM AUTHOR]

To solve the problem of the poor enhancement effect of traditional deep learning-based speech enhancement algorithms in low signal-to-noise ratio (SNR) scenarios, this paper proposes a method combining front-end processing Multi-Resolution Cochleagram(FP-MRCG) and skip connections deep neural network (Skip-DNN). This method uses FP-MRCG speech features to train Skip-DNN, and estimates the ideal ratio mask, filters out the background noise of the noisy speech to obtain the enhanced speech features, and obtains enhanced speech by phase reconstruction. The result shows that when the SNR is 0dB, using FP-MRCG as Skip-DNN's input, the average perceptual evaluation of speech quality (PESQ) of enhanced speech is 2.5283, and the average short-term objective intelligibility (STOI) is 0.8825, which is 3 % and 1.7 % higher than MRCG, respectively. Besides, when using FP-MRCG as the input of DNN, Skip-DNN and convolutional neural network (CNN), Skip-DNN has a higher evaluation score in a low SNR environment, and CNN has a higher evaluation score in a high SNR environment. However, the training time for the CNN is twice as long as that for the Skip-DNN. Hence, it can be concluded that Skip-DNN performs better in speech enhancement than the other two networks. [ABSTRACT FROM AUTHOR]

The coronavirus disease 2019 (COVID-19) pandemic has been a globally dangerous crisis that causes an increasingly high death rate. Applying machine learning to the computed-tomography (CT)-based COVID-19 diagnosis is essential and attracts the attention of the research community. This paper introduces an approach for simultaneously identifying COVID-19 disease and segmenting its manifestations on lung images. The proposed method is an asymmetric U-Net-like model improved with skip connections. The experiment was conducted on a light-weighted feature extractor called CRNet with a feature enhancement technique called atrous spatial pyramid pooling. Classifying between COVID-19 and non-COVID-19 cases recorded the highest mean scores of 97.1, 94.4, and 97.0% for accuracy, dice similarity coefficient (DSC) and F1 score, respectively. Alternatively, the respective highest mean scores of the classification between COVID-19 and community-acquired pneumonia were 99.89, 99.79, and 99.97%. The lesion segmentation performance was with the highest mean of 99.6 and 84.7% for, respectively, accuracy and DSC. [ABSTRACT FROM AUTHOR]

Copyright of Chinese Journal of Liquid Crystal & Displays is the property of Chinese Journal of Liquid Crystal & Displays and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Perilla seed meal (PSM), a byproduct of oil extraction from Perilla frutescens L. seeds, is rich in protein (24.26–42.85%) and holds potential as an economical and sustainable animal feed. Traditional methods for assessing protein content are labor-intensive and costly. This study explores Near-Infrared Reflectance Spectroscopy (NIRS) for the rapid, precise, and non-destructive determination of PSM protein content in 126 samples. We developed and evaluated Modified Partial Least Squares (MPLS) regression and deep learning (DL) models, including 1D-CNN (Convolutional Neural Network), LSTM Long Short-Term Memory), and hybrid architectures incorporating skip connections, inception modules, and spectral derivatives. Model performance was validated externally using parameters such as RSQ external (R-squared), bias, SEP(C) (Standard Error of Prediction), RPD (Residual Prediction Deviation), slope, SD (Standard Deviation), p-value (≥0.05), and the correlation between reference and predicted values. The 1D CNN-LSTM-Inception derivative 1 model achieved the best performance (RPD: 8.0, RSQ external : 0.98), followed by the MPLS-based model (RPD: 4.88, RSQ external : 0.96) and the 1D CNN derivative 1 model (RPD: 3.07, RSQ external : 0.96). These models provide a reliable and advanced technology for the non-destructive screening of PSM protein content, thus aiding in the rapid identification and selection of superior perilla chemotypes from varied backgrounds. [ABSTRACT FROM AUTHOR]

In this work we introduce a biologically inspired long-range skip connection for the UNet architecture that relies on the perceptual illusion of hybrid images, being images that simultaneously encode two images. The fusion of early encoder features with deeper decoder ones allows UNet models to produce finer-grained dense predictions. While proven in segmentation tasks, the network’s benefits are down-weighted for dense regression tasks as these long-range skip connections additionally result in texture transfer artifacts. Specifically for depth estimation, this hurts smoothness and introduces false positive edges which are detrimental to the task due to the depth maps’ piece-wise smooth nature. The proposed HybridSkip connections show improved performance in balancing the trade-off between edge preservation, and the minimization of texture transfer artifacts that hurt smoothness. This is achieved by the proper and balanced exchange of information that HybridSkip connections offer between the high and low frequency, encoder and decoder features, respectively. The code and models will be made available in the project page.

SIDA-GAN is a set of GAN based deep learning models that can be employed for predicting the depth map for a given single input RGB image. This article explores the research gap present in the state of the art which lacks GAN based depth predictors implemented using Depthwise Separable Convolution (DSC). DSC reduces the number of computations in comparison with standard convolution for the same convolution task, therefore making the model computationally efficient. The adversarial loss applied by the GAN model make the predicted depth image globally coherent while loss functions like L1, SSIM fine tunes the local feature predictions leading to enhanced prediction quality. In this proposal, the depth predictor (SIDA-GAN generator) is an encoder-decoder design featuring DSC and skip connections. SIDA-GAN has two GAN variants namely conditional GAN (cGAN), patch GAN (pGAN) and two generator implementations based on MobileNetV1, MobileNetV2 leading to four different flavours. Those are SIDA-cGAN (MobileNetV1 based generator), SIDA-cGAN (MobileNetV2 based generator), SIDA-pGAN (MobileNetV1 based generator), SIDA-pGAN (MobileNetV2 based generator). The proposed models are trained with different combinations of SSIM, L1, depth-smoothness and adversarial losses using NYU-Depth V2 (Nathan Silberman et al., 2012) dataset. SIDA-GAN has achieved reduction in trainable parameter count up to 76% and performance improvement in Average Relative Error (8%–25%), Root Mean Square Error (3%–27%), threshold accuracy (2%–6%) when compared with baseline papers.

In recent years, convolutional neural-network-based crack segmentation methods have performed excellently. However, existing crack segmentation methods still suffer from background noise interference, such as dirt patches and pitting, as well as the imprecise segmentation of fine-grained spatial structures. This is mainly due to the fact that convolutional neural networks dilute low-level spatial information in the process of extracting deep semantic features, and the network cannot obtain accurate context awareness because of the limitation of the actual receptive field size. To address these problems, an encoder–decoder crack segmentation network based on multi-scale contextual information enhancement is proposed. First, a new architecture of skip connection is proposed, enabling the network to obtain refined crack segmentation results; then, a contextual feature enhancement module is designed to make the network more effective at distinguishing between cracks and background noise; finally, the deformable convolution is introduced into the encoder network to further enhance its ability to extract the diverse morphological features of cracks by adaptively adjusting the sampling area and the receptive field size. Experiments show that the proposed method is effective in crack segmentation and outperforms mainstream segmentation networks such as DeepLab V3+ and UNet++. [ABSTRACT FROM AUTHOR]

Among researchers using traditional and new machine learning and deep learning techniques, 2D medical image segmentation models are popular. Additionally, 3D volumetric data recently became more accessible, as a result of the high number of studies conducted in recent years regarding the creation of 3D volumes. Using these 3D data, researchers have begun conducting research on creating 3D segmentation models, such as brain tumor segmentation and classification. Since a higher number of crucial features can be extracted using 3D data than 2D data, 3D brain tumor detection models have increased in popularity among researchers. Until now, various significant research works have focused on the 3D version of the U-Net and other popular models, such as 3D U-Net and V-Net, while doing superior research works. In this study, we used 3D brain image data and created a new architecture based on a 3D U-Net model that uses multiple skip connections with cost-efficient pretrained 3D MobileNetV2 blocks and attention modules. These pretrained MobileNetV2 blocks assist our architecture by providing smaller parameters to maintain operable model size in terms of our computational capability and help the model to converge faster. We added additional skip connections between the encoder and decoder blocks to ease the exchange of extracted features between the two blocks, which resulted in the maximum use of the features. We also used attention modules to filter out irrelevant features coming through the skip connections and, thus, preserved more computational power while achieving improved accuracy. [ABSTRACT FROM AUTHOR]

Acute lymphoblastic leukemia (ALL) is a life-threatening hematological malignancy that requires early and accurate diagnosis for effective treatment. However, the manual diagnosis of ALL is time-consuming and can delay critical treatment decisions. To address this challenge, researchers have turned to advanced technologies such as deep learning (DL) models. These models leverage the power of artificial intelligence to analyze complex patterns and features in medical images and data, enabling faster and more accurate diagnosis of ALL. However, the existing DL-based ALL diagnosis suffers from various challenges, such as computational complexity, sensitivity to hyperparameters, and difficulties with noisy or low-quality input images. To address these issues, in this paper, we propose a novel Deep Skip Connections-Based Dense Network (DSCNet) tailored for ALL diagnosis using peripheral blood smear images. The DSCNet architecture integrates skip connections, custom image filtering, Kullback–Leibler (KL) divergence loss, and dropout regularization to enhance its performance and generalization abilities. DSCNet leverages skip connections to address the vanishing gradient problem and capture long-range dependencies, while custom image filtering enhances relevant features in the input data. KL divergence loss serves as the optimization objective, enabling accurate predictions. Dropout regularization is employed to prevent overfitting during training, promoting robust feature representations. The experiments conducted on an augmented dataset for ALL highlight the effectiveness of DSCNet. The proposed DSCNet outperforms competing methods, showcasing significant enhancements in accuracy, sensitivity, specificity, F-score, and area under the curve (AUC), achieving increases of 1.25%, 1.32%, 1.12%, 1.24%, and 1.23%, respectively. The proposed approach demonstrates the potential of DSCNet as an effective tool for early and accurate ALL diagnosis, with potential applications in clinical settings to improve patient outcomes and advance leukemia detection research.

Some recent studies show that filters in convolutional neural networks (CNNs) have low color selectivity in datasets of natural scenes such as Imagenet. CNNs, bio-inspired by the visual cortex, are characterized by their hierarchical learning structure which appears to gradually transform the representation space. Inspired by the direct connection between the LGN and V4, which allows V4 to handle low-level information closer to the trichromatic input in addition to processed information that comes from V2/V3, we propose the addition of a long skip connection (LSC) between the first and last blocks of the feature extraction stage to allow deeper parts of the network to receive information from shallower layers. This type of connection improves classification accuracy by combining simple-visual and complex-abstract features to create more color-selective ones. We have applied this strategy to classic CNN architectures and quantitatively and qualitatively analyzed the improvement in accuracy while focusing on color selectivity. The results show that, in general, skip connections improve accuracy, but LSC improves it even more and enhances the color selectivity of the original CNN architectures. As a side result, we propose a new color representation procedure for organizing and filtering feature maps, making their visualization more manageable for qualitative color selectivity analysis.

In recent years,deep learning has been widely used in image super-resolution reconstruction.To solve the problems of inadequate feature extraction,loss of details and gradient disappearance in super-resolution reconstruction methods based on deep learning,a deep recursive residual neural network model based on channel attention is proposed for single image super-resolution reconstruction.The proposed model constructs a simple recursive residual network structure by residual nested networks and jump connections to deepen the network and speed up its convergence while avoiding network degradation and gradient problems.An attention mechanism is introduced into the feature extraction part to improve the discriminant learning ability of the network for more accurate and more effective extraction of deep residual features,which is combined with the subsequent reconstruction network with parallel mapping structure to ensure the final accurate reconstruction.Quantitative and qualitative assessments are performed on benchmark dataset Set5,Set14,B100 and Urban100 at the magnification of 2,3 and 4 times by comparison with the mainstream methods.Experimental results show that the objective index values of the proposed method increase significantly compared to the comparative methods on all four test data sets.Among them,compared with the interpolation method and the SRCNN algorithm,the average PSNR improves 3.965dB and 1.56dB,3.19dB and 1.42dB,2.79dB and 1.32dB,respectively,at the magnification of 2,3 and 4 times.Visual effects show that the proposed method can recover image details better.

Automated magnetic resonance imaging (MRI) segmentation plays an important role for the diagnosis and evaluation of brain tumors. However, it involves many challenges, including large variations in the shape and types of brain tumors, especially small ones that are difficult to diagnose early, and the complex border patterns that lead to missing data and low segmentation accuracy, putting specialists under time pressure. In this paper, we propose a new, lightweight network based on Unet by integrating skip connections with depthwise convolutional layers. This complementary technique allows us to improve the performance of the Unet model in explaining and analyzing medical images in a comprehensive and effective manner by transferring information from advanced layers to lower layers with greater quality. It helps avoid information loss and contributes to improving the network's ability to extract features better. Our suggested approach produced superior results compared to other Unet-like networks, as evidenced by the following metrics: accuracy of 99.82%, Intersection over union (IoU) of 83.34%, and Dice Similarity Coefficient (DSC) of 90.83%. Experimental findings demonstrated that the skip connections reconstructed in depthwise Convolution Bottleneck (DCB-Unet) integrated multi-scale contextual semantic information. In contrast to the majority of contemporary medical image segmentation models, our suggested approach more precisely recognizes organs, lesions, and segments. [ABSTRACT FROM AUTHOR]

Neural networks, while powerful, often lack interpretability. Physics-Informed Neural Networks (PINNs) address this limitation by incorporating physics laws into the loss function, making them applicable to solving Ordinary Differential Equations (ODEs) and Partial Differential Equations (PDEs). The recently introduced PINC framework extends PINNs to control applications, allowing for open-ended long-range prediction and control of dynamic systems. In this work, we enhance PINC for modeling highly nonlinear systems such as gas-lifted oil wells. By introducing skip connections in the PINC network and refining certain terms in the ODE, we achieve more accurate gradients during training, resulting in an effective modeling process for the oil well system. Our proposed improved PINC demonstrates superior performance, reducing the validation prediction error by an average of 67% in the oil well application and significantly enhancing gradient flow through the network layers, increasing its magnitude by four orders of magnitude compared to the original PINC. Furthermore, experiments showcase the efficacy of Model Predictive Control (MPC) in regulating the bottom-hole pressure of the oil well using the improved PINC model, even in the presence of noisy measurements. • Physics-informed neural nets for control (PINC) of complex nonlinear plants. • Skip connections in PINC markedly improves gradient flow in network layers. • Improved PINC significantly reduces validation error by 67% in oil well application. • Applying MPC with PINC precisely controls oil well bottom-hole pressure. • Enhanced PINC proves robust against noisy measurements in oil well control. [ABSTRACT FROM AUTHOR]

Underwater images get degraded for a variety of naturally occurring attributes like haze, suspended particles, light scattering and water types. The primary cause of the degradation is underwater light attenuation that varies with wavelength, unlike the uniform attenuation that occurs in-air. In this paper, we propose an end-to-end deep convolutional neural network architecture to restore the underwater images and improve their visual perception. The encoder learns to encode the degraded image to a lower-dimensional feature map, while the decoder learns to restore the image to a degradation-free form. This is achieved due to the utilization of symmetric skip connections between the encoder–decoder blocks for the propagation of feature maps to improve the sharpness of the restored image and prevent the loss of details caused by the convolutions. We exhaustively evaluate the performance of our network both qualitatively and quantitatively on standard datasets, and the effectiveness of our network is demonstrated with existing methods of underwater image restoration and enhancement techniques. [ABSTRACT FROM AUTHOR]

In this paper, a joint restoration convolutional neural network (JRCNN) is proposed to produce a visually pleasing super resolution (SR) image from a single low-quality (LQ) image. The LQ image is a low resolution (LR) image with ringing, blocking and blurring artifacts arising due to compression. JRCNN consists of three deep dense residual blocks (DRB). Each DRB comprises of parallel convolutional layers with cross residual connections. The representational power of JRCNN is improved by depth-wise concatenation of feature representations from each of the DRBs. Moreover, these connections mitigate the problem of vanishing of gradients. Different from the previous networks, JRCNN exploits the contextual information directly in the LR image space without using any interpolation. This strategy improves the training efficiency of the network. The exhaustive experimentation on different datasets show that the proposed JRCNN produces state-of-the-art performance. Furthermore, ablation experiments are performed to assess the effectiveness of JRCNN. In addition, individual experiments are conducted for SR and compression artifact removal on benchmark datasets. [ABSTRACT FROM AUTHOR]

BACKGROUND: The noise in magnetic resonance (MR) images causes severe issues for medical diagnosis purposes. OBJECTIVE: In this paper, we propose a discriminative learning based convolutional neural network denoiser to denoise the MR image data contaminated with noise. METHODS: The proposed method incorporates the use of depthwise separable convolution along with local response normalization with modified hyperparameters and internal skip connections to denoise the contaminated MR images. Moreover, the addition of parametric RELU instead of normal conventional RELU in our proposed architecture gives more stable and fine results. The denoised images were further segmented to test the appropriateness of the results. The network is trained on one dataset and tested on other dataset produces remarkably good results. RESULTS: Our proposed network was used to denoise the images of different noise levels, and it yields better performance as compared with various networks. The SSIM and PSNR showed an average improvement of (7.2 ± 0.002) % and (8.5 ± 0.25) % respectively when tested on different datasets without retaining the network. An improvement of 5% and 6% was achieved in the values of mean intersection over union (mIoU) and BF score when the denoised images were segmented for testing the relevancy in biomedical imaging applications. The statistical test suggests that the obtained results are statistically significant as p < 0.05. CONCLUSION: The denoised images obtained are more clinically suitable for medical image diagnosis purposes, as depicted by the evaluation parameters. Further, external clinical validation was performed by an experienced radiologist for testing the validation of the resulting images. [ABSTRACT FROM AUTHOR]

Convolutional neural networks hold state-of-the-art results for image classification, and many neural architecture search algorithms have been proposed to discover high performance convolutional neural networks. However, the use of neural architecture search for the discovery of skip-connection structures, an important element in modern convolutional neural networks, is limited within the literature. Furthermore, while many neural architecture search algorithms utilize performance estimation techniques to reduce computation time, empirical evaluations of these performance estimation techniques remain limited. This work focuses on utilizing evolutionary neural architecture search to examine the search space of networks, which follow a fundamental DenseNet structure, but have no fixed skip connections. In particular, a genetic algorithm is designed, which searches the space consisting of all networks between a standard feedforward network and the corresponding DenseNet. To design the algorithm, lower fidelity performance estimation of this class of networks is examined and presented. The final algorithm finds networks that are more accurate than DenseNets on CIFAR10 and CIFAR100, and have fewer trainable parameters. The structures found by the algorithm are examined to shed light on the importance of different types of skip-connection structures in convolutional neural networks, including the discovery of a simple skip-connection removal, which improves DenseNet performance on CIFAR10. [ABSTRACT FROM AUTHOR]