Convolution Kernel Research Articles

Metasurface-Generated Large and Arbitrary Analog Convolution Kernels for Accelerated Machine Vision

Metasurface-Generated Large and Arbitrary Analog Convolution Kernels for Accelerated Machine Vision

Deep learning and time series signal processing for bending detection in mining environment using optical fiber sensor

Optical fiber sensors are widely used to monitor environmental disturbances that may trigger or affect mining landslides, a complex phenomenon that involves various factors and can cause severe damage. In this study, we propose a method to detect the curvature of a bent optical fiber, which affects the optical signals traveling through it, using deep learning and time series signal processing. We use a Mach-Zehnder interferometer (MZI) sensor and a single-mode fiber (SMF) cable to capture the interference signal caused by different bending angles. We apply a transformation random convolutional kernel (TRANCE) to reduce the signal and extract its mean and standard deviation as features. We train a dense neural network (DNN) with three hidden layers to classify the bending events into four categories: low, medium, high, and extreme. We evaluate our method on a dataset of 400 samples and achieve an average accuracy of 99% in the confusion matrix and the fifty times tenfold validation experiments. Our method outperforms other methods that use multimode fiber (MMF), speckle images, and convolutional neural networks (CNNs) in terms of accuracy, speed, and generalization. Our method also has the advantages of using SMF, which enables integrated sensing and communication, and using 1D signals, which are easier to process than 2D images. Our method is suitable for high-speed signals containing information in dangerous environments.

Deep learning model for detection acute cardiogenic pulmonary edema in cases of preeclampsia

<span lang="EN-US">The physiological changes during the pregnancy period increase the risk of developing pulmonary edema and acute respiratory failure. This condition falls under critical medical emergencies associated with maternal mortality. This study utilized a convolutional neural networks (CNN) architectural model employing chest Xray dataset images. CNN utilizes the convolution process by moving a convolutional kernel of a certain size across an image, allowing the computer to derive new representative information from the multiplication of portions of the image with the utilized filter.</span><span lang="EN-US">To simplify, the vanishing gradient issue occurs when information dissipates before reaching its destination due to the lengthy path between input and output layers. This study was developed model for detection acute cardiogenic pulmonary Edema in pre-eclampsia cases using chest Xray images, implemented using PyTorch, Keras, and MxNet. The validated model achieved its optimum with accuracy 90.65% and binary cross-entropy loss (BCELoss) value of 0.4538. It exhibited an improved sensitivity value of 83.514% using a 5% dataset and a specificity value of 57.273%. This indicates an increase in sensitivity value by 83.514% using a 5% data set and a specificity value of 57.273%. The research results demonstrate an improvement in accuracy compared to several similar studies that also utilized CNN models.</span>

Eliminating electromagnetic interference for RF shielding-free MRI via k-space convolution: Insights from MR parallel imaging advances.

Recent advances in ultra-low field MRI have attracted attention from both academic and industrial MR communities for its potential in democratizing MRI applications. One of the most striking features on those advances is shielding-free imaging by actively sensing and eliminating the electromagnetic interference (EMI). In this study, we review the analytical approaches for EMI estimation/elimination, and investigate their theoretical basis and relations with parallel imaging reconstruction. We provide further understanding of the existing approaches, formulating EMI estimation as convolution in k-space or multiplication in spectrum-space. We further propose to use tailored convolutional kernel to adaptively fit the varying EMI coupling across the acquisition window. These methods were evaluated with both simulation study and human brain imaging. The results show that using tailored convolutional kernel can achieve more robust performance against system and acquisition imperfections.

Reducing overfitting in vehicle recognition by decorrelated sparse representation regularisation

AbstractMost state‐of‐the‐art vehicle recognition methods benefit from the excellent feature extraction capabilities of convolutional neural networks (CNNs), which allow the models to perform well on the intra‐dataset. However, they often show poor generalisation when facing cross‐datasets due to the overfitting problem. For this issue, numerous studies have shown that models do not generalise well in new scenarios due to the high correlation between the representations in CNNs. Furthermore, over‐parameterised CNNs have a large number of redundant representations. Therefore, we propose a novel Decorrelated Sparse Representation (DSR) regularisation. (1) It tries to minimise the correlation between feature maps to obtain decorrelated representations. (2) It forces the convolution kernels to extract meaningful features by allowing the sparse kernels to have additional optimisation. The DSR regularisation encourages diverse representations to reduce overfitting. Meanwhile, DSR can be applied to a wide range of vehicle recognition methods based on CNNs, and it does not require additional computation in the testing phase. In the experiments, DSR performs better than the original model on the intra‐dataset and cross‐dataset. Through ablation analysis, we find that DSR can drive the model to focus on the essential differences among all kinds of vehicles.

Enhancing low-light images using Sakaguchi type function and Gegenbauer polynomial

Enhancing low-light images is crucial for various applications in computer vision, yet current approaches often fall short in balancing image quality and detail preservation. This study introduces a novel method designed to enhance low-light images by applying advanced mathematical techniques from geometric function theory. Specifically, we employ Sakaguchi-type class functions, subordinated with the Gegenbeur polynomial, to derive coefficient estimations. These estimations are then used in convolution kernels to produce enhanced image versions. The method was tested on the LOw-Light dataset (LOL), containing challenging low-light images with noise and artifacts. Our approach’s effectiveness is validated through quantitative metrics, including PSNR and SSIM, as well as visual comparisons. The results demonstrate significant improvements over existing state-of-the-art methods, offering better visibility and detail retention. This method holds promise for enhancing images in critical fields such as surveillance and medical imaging.

Lie group convolution neural networks with scale-rotation equivariance

The weight-sharing mechanism of convolutional kernels ensures the translation equivariance of convolutional neural networks (CNNs) but not scale and rotation equivariance. This study proposes a SIM(2) Lie group-CNN, which can simultaneously keep scale, rotation, and translation equivariance for image classification tasks. The SIM(2) Lie group-CNN includes a lifting module, a series of group convolution modules, a global pooling layer, and a classification layer. The lifting module transfers the input image from Euclidean space to Lie group space, and the group convolution is parameterized through a fully connected network using the Lie Algebra coefficients of Lie group elements as inputs to achieve scale and rotation equivariance. It is worth noting that the mapping relationship between SIM(2) and its Lie Algebra and the distance measure of SIM(2) are defined explicitly in this paper, thus solving the problem of the metric of features on the space of SIM(2) Lie group, which contrasts with other Lie groups characterized by a single element, such as SO(2). The scale-rotation equivariance of Lie group-CNN is verified, and the best recognition accuracy is achieved on three categories of image datasets. Consequently, the SIM(2) Lie group-CNN can successfully extract geometric features and perform equivariant recognition on images with rotation and scale transformations.

Kernel Determination Problem in the Third Order 1D Moore-Gibson-Thompson Equation with Memory

In this study, we address the inverse problem of determining the convolution kernel function in the third-order Moore--Gibson--Thompson (MGT) equation, which is commonly used to model fluid motion with memory effects. Specifically, we focus on the determination of the unknown kernel, which governs the memory term in the equation. Initially, we employ the Fourier spectral method to solve the direct initial-boundary-value problem for a non-homogeneous MGT equation that with the memory term. The Fourier spectral method allows us to leverage the problem's inherent linearity and spatial homogeneity, leading to an efficient and explicit construction of the solution. The direct problem is analyzed under appropriate initial and boundary conditions, which are carefully specified to ensure mathematical consistency. To solve the inverse problem, we introduce an additional condition-typically a form of observational data such as at certain points-which provides the necessary constraints for determining the kernel. We prove local existence and uniqueness theorems for solution of the problem.

Traveling wave solutions of a diffusive predator-prey system with Holling II type functional response

In this paper, we discuss a three-dimensional diffusive predator-prey system with nonlocal terms and Holling II type functional response. According to the relationship between traveling wave and heteroclinic orbit, the predator-prey system is transformed into the singularly perturbed system. Based on the method of the geometric singular perturbation theory, we construct a locally invariant manifold to obtain the traveling wave solutions with nonlocal delay convolution kernel.

Automated Identification and Segmentation of H i Sources in CRAFTS Using Deep Learning Method

Abstract Identifying neutral hydrogen (H i) galaxies from observational data is a significant challenge in H i galaxy surveys. With the advancement of observational technology, especially with the advent of large-scale telescope projects such as FAST and SKA, the significant increase in data volume presents new challenges for the efficiency and accuracy of data processing.To address this challenge, in this study, we present a machine learning-based method for extracting H i sources from the three-dimensional (3D) spectral data obtained from the Commensal Radio Astronomy FAST Survey (CRAFTS). We have carefully assembled a specialized dataset, HISF, rich in H i sources, specifically designed to enhance the detection process. Our model, Unet-LK, utilizes the advanced 3D-Unet segmentation architecture and employs an elongated convolution kernel to effectively capture the intricate structures of H i sources. This strategy ensures a reliable identification and segmentation of H i sources, achieving notable performance metrics with a recall rate of 91.6% and an accuracy of 95.7%. These results substantiate the robustness of our dataset and the effectiveness of our proposed network architecture in the precise identification of H i sources. Our code and dataset is publicly available at https://github.com/fishszh/HISF.

Sa-SNN: spiking attention neural network for image classification

Spiking neural networks (SNNs) are known as third generation neural networks due to their energy efficient and low power consumption. SNNs have received a lot of attention due to their biological plausibility. SNNs are closer to the way biological neural systems work by simulating the transmission of information through discrete spiking signals between neurons. Influenced by the great potential shown by the attention mechanism in convolutional neural networks, Therefore, we propose a Spiking Attention Neural Network (Sa-SNN). The network includes a novel Spiking-Efficient Channel Attention (SECA) module that adopts a local cross-channel interaction strategy without dimensionality reduction, which can be achieved by one-dimensional convolution. It is implemented by convolution, which involves a small number of model parameters but provides a significant performance improvement for the network. The design of local inter-channel interactions through adaptive convolutional kernel sizes, rather than global dependencies, allows the network to focus more on the selection of important features, reduces the impact of redundant features, and improves the network’s recognition and generalisation capabilities. To investigate the effect of this structure on the network, we conducted a series of experiments. Experimental results show that Sa-SNN can perform image classification tasks more accurately. Our network achieved 99.61%, 99.61%, 94.13%, and 99.63% on the MNIST, Fashion-MNIST, N-MNIST datasets, respectively, and Sa-SNN performed well in terms of accuracy compared with mainstream SNNs.

LVGG-IE: A Novel Lightweight VGG-Based Image Encryption Scheme

Image security faces increasing challenges with the widespread application of computer science and artificial intelligence. Although chaotic systems are employed to encrypt images and prevent unauthorized access or tampering, the degradation that occurs during the binarization process in chaotic systems reduces security. The chaos- and DNA-based image encryption schemes increases its complexity, while the integration of deep learning with image encryption is still in its infancy and has several shortcomings. An image encryption scheme with high security and efficiency is required for the protection of the image. To address these problems, we propose a novel image encryption scheme based on the lightweight VGG (LVGG), referred to as LVGG-IE. In this work, we design an LVGG network with fewer layers while maintaining a high capacity for feature capture. This network is used to generate a key seed, which is then employed to transform the plaintext image into part of the initial value of a chaotic system, ensuring that the chaos-based key generator correlates with the plaintext image. A dynamic substitution box (S-box) is also designed and used to scramble the randomly shuffled plaintext image. Additionally, a single-connected (SC) layer is combined with a convolution layer from VGG to encrypt the image, where the SC layer is dynamically constructed by the secret key and the convolution kernel is set to . The encryption efficiency is simulated, and the security is analyzed. The results show that the correlation coefficient between adjacent pixels in the proposed scheme achieves . The NPCR exceeds 0.9958, and the UACI falls within the theoretical value with a significance level of 0.05. The encryption quality, the security of the dynamic S-box and the SC layer, and the efficiency are tested. The result shows that the proposed image encryption scheme demonstrates high security, efficiency, and robustness, making it effective for image security in various applications.

Multi-architecture optimization of pipeline inner wall defect detection algorithm based on YOLOv8

As traditional computer vision technology struggles to meet the demands for accurate detection of modern supply pipes. A detection model based on an improved YOLOv8 network has been proposed. First, the Inverted Residual Mobile Block (iRMB) is integrated into the backbone network. This effectively enhances the feature representation capability for extracting image defects. Next, a shift-wise operator is introduced to simulate the effects of using a large convolutional kernel at a lower computational cost while improving performance. Finally, GSConv replaces the Conv layer in the neck network, and the VoVGSCSP module substitutes the C2f module in the neck network to further enhance the algorithm’s detection accuracy for corroded regions. Experimental results demonstrate that the mean Average Precision (mAP) of the improved algorithm reaches 95.0 % on the dataset of pipeline inner wall corrosion defects. This provides an accurate method for the intelligent identification of corrosion defects on the pipeline dataset.

ARSOD-YOLO: Enhancing Small Target Detection for Remote Sensing Images

Remote sensing images play a vital role in domains including environmental monitoring, agriculture, and autonomous driving. However, the detection of targets in remote sensing images remains a challenging task. This study introduces innovative methods to enhance feature extraction, feature fusion, and model optimization. The Adaptive Selective Feature Enhancement Module (AFEM) dynamically adjusts feature weights using GhostModule and sigmoid functions, thereby enhancing the accuracy of small target detection. Moreover, the Adaptive Multi-scale Convolution Kernel Feature Fusion Module (AKSFFM) enhances feature fusion through multi-scale convolution operations and attention weight learning mechanisms. Moreover, our proposed ARSOD-YOLO optimized the network architecture, component modules, and loss functions based on YOLOv8, enhancing outstanding small target detection capabilities while preserving model efficiency. We conducted experiments on the VEDAI and AI-TOD datasets, showcasing the excellent performance of ARSOD-YOLO. Our algorithm achieved an mAP50 of 74.3% on the VEDAI dataset, surpassing the YOLOv8 baseline by 3.1%. Similarly, on the AI-TOD dataset, the mAP50 reached 47.8%, exceeding the baseline network by 6.1%.

Human key point detection method based on enhanced receptive field and transformer

The existing key point detection network models have complex structures, so it is difficult to deploy on edge devices. Meanwhile, the convolution is localized and limited by the size of convolution kernels, which cannot effectively capture long-range dependencies. To address this problem, this paper introduces a lightweight Convolutional Transformer network (LHFormer Net) for human pose estimation. Considering that the sampling area of convolution kernels for different dimensional feature maps is fixed and the contextual information is singular, the enhanced receptive field block is designed to extract richer feature information and reduce information loss in feature maps. Based on the global modeling features of Transformer encoder, convolutional position encoding and multi-head self-attention are used to capture the spatial constraint relationship between key points in the deep feature extraction. Finally, a lightweight deconvolution module is used to generate higher resolution features to achieve multi-resolution supervision, which can effectively solve the problem of scale variation in pose estimation, more accurately locate key points of small and medium-sized people, ad further improve the detection accuracy of the network. Compared with other networks, the experimental results on the open access datasets COCO2017 and MPII show that the proposed network achieves a good balance between model complexity and detection performance.

Online learning discriminative sparse convolution networks for robust UAV object tracking

Despite the remarkable empirical success for UAV object tracking, current convolutional networks usually have three unavoidable limitations: (1) The feature maps produced by convolutional layers are difficult to interpret. (2) The network needs to be trained offline on a large-scale auxiliary training set, resulting in the feature extraction ability of the trained network depending on the categories of the training set. (3) The performance of networks suffers from sensitivity to hyper-parameters (such as learning rate and weight decay) when the network needs online fine-tuning. To overcome the three limitations, this paper proposes a Discriminative Sparse Convolutional Network (DSCN) that exhibits good layer-wise interpretability and can be trained online without requiring any auxiliary training data. By imposing sparsity constraints on the convolutional kernels, DSCN furnishes the convolution layer with an explicit data meaning, thus enhancing the interpretability of the feature maps. These convolutional kernels are directly learned online from image blocks, which eliminates the offline training process on auxiliary data sets. Moreover, a simple yet effective online tuning method with few hyper-parameters is proposed to fine-tune fully connected layers online. We have successfully applied DSCN to UAV object tracking and conducted extensive experiments on six mainstream UAV datasets. The experimental results demonstrate that our method performs favorably against several state-of-the-art tracking algorithms in terms of tracking accuracy and robustness.

Pyramid boundary attention network for breast lesion segmentation in ultrasound images

Breast cancer is one of the most common causes of death among women worldwide and poses a major threat to women’s health. Currently, ultrasound examination is widely used for early detection of breast cancer, and accurate segmentation of breast lesions from ultrasound images is beneficial for early diagnosis of breast cancer. In this study, a pyramid boundary attention network was developed to automatically and stably segment breast lesions. In addition, a pyramid squeeze attention module was introduced into U-Net to increase the network’s receptive field and establish global attention by using convolutional kernels of different sizes, thereby improving the network’s detection and segmentation capabilities for tumors of different sizes. Additionally, to address the problem of fuzzy boundaries of breast tumor in ultrasound images, a boundary branch network was designed that used the contextual local attention module to extract the feature maps of different dimensions from the encoder and integrated local spatial and high-level semantic context information. The boundary fusion module was used to extract global and local information of tumor boundaries to obtain the boundary map of tumors, thereby refining the boundary segmentation of tumors and improving the segmentation accuracy of tumors. The proposed method was validated on two public datasets, BUSI and Dataset B, with dice coefficients of 78.98% and 84.5%, respectively. Compared with other medical image segmentation methods and the latest semantic segmentation methods, the proposed method showed superior performance in segmenting breast lesion in ultrasound images.

Adaptive Kernel Convolutional Stereo Matching Recurrent Network.

For binocular stereo matching techniques, the most advanced method currently is using an iterative structure based on GRUs. Methods in this class have shown high performance on both high-resolution images and standard benchmarks. However, simply replacing cost aggregation with a GRU iterative method leads to the original cost volume for disparity calculation lacking non-local geometric and contextual information. Based on this, this paper proposes a new GRU iteration-based adaptive kernel convolution deep recurrent network architecture for stereo matching. This paper proposes a kernel convolution-based adaptive multi-scale pyramid pooling (KAP) module that fully considers the spatial correlation between pixels and adds new matching attention (MAR) to refine the matching cost volume before inputting it into the iterative network for iterative updates, enhancing the pixel-level representation ability of the image and improving the overall generalization ability of the network. At present, the AKC-Stereo network proposed in this paper has a higher improvement than the basic network. On the Sceneflow dataset, the EPE of AKC-Stereo reaches 0.45, which is 0.02 higher than the basic network. On the KITTI 2015 dataset, the AKC-Stereo network outperforms the base network by 5.6% on the D1-all metric.

Low-frequency spectral graph convolution networks with one-hop connections information for personalized tag recommendation

Graph neural networks (GNNs) have gained prominence as an effective technique for representation learning and have found wide application in tag recommendation tasks. Existing approaches aim to encode the hidden collaborative information among entities into embedding representations by propagating node information between connected nodes. However, in sparse observable graph structures, a significant number of connections are missing, leading to incomplete and biased propagation. To address these issues, we propose a novel model called Low-frequency Spectral Graph Convolution Networks with one-hop connections information for Personalized Tag Recommendation (LSGCNT). This model utilizes graph convolution in the spectral domain and incorporates a graph structure comprising two bipartite graphs, the user–tag interaction graph and the item–tag interaction graph. Our model aims to reduce information loss caused by propagation by utilizing graph convolution networks with trainable convolution kernels to recover preference information. In order to preserve useful low-frequency signals, we couple graph convolution with low-pass filters in the frequency domain. Through reconstructing the true rating tensor and ranking the tag scores within the tensor, we can achieve top-N recommendations. Furthermore, to preserve the one-hop connection information of the bipartite graphs, we treat the observed two bipartite graphs as two homogeneous graphs, where both users and tags contribute to the convolution of a node in the user–tag graph, and both items and tags contribute to the convolution of a node in the item–tag graph. Lastly, we analyze the impact of different internal components, pooling methods, parameter choices, and prediction approaches of LSGCNT on recommendation performance. Experimental results on two real-world datasets demonstrate that LSGCNT achieves superior recommendation performance compared with eight other state-of-the-art recommendation models.

TSdetector: Temporal–Spatial self-correction collaborative learning for colonoscopy video detection

CNN-based object detection models that strike a balance between performance and speed have been gradually used in polyp detection tasks. Nevertheless, accurately locating polyps within complex colonoscopy video scenes remains challenging since existing methods ignore two key issues: intra-sequence distribution heterogeneity and precision-confidence discrepancy. To address these challenges, we propose a novel Temporal–Spatial self-correction detector (TSdetector), which first integrates temporal-level consistency learning and spatial-level reliability learning to detect objects continuously. Technically, we first propose a global temporal-aware convolution, assembling the preceding information to dynamically guide the current convolution kernel to focus on global features between sequences. In addition, we designed a hierarchical queue integration mechanism to combine multi-temporal features through a progressive accumulation manner, fully leveraging contextual consistency information together with retaining long-sequence-dependency features. Meanwhile, at the spatial level, we advance a position-aware clustering to explore the spatial relationships among candidate boxes for recalibrating prediction confidence adaptively, thus eliminating redundant bounding boxes efficiently. The experimental results on three publicly available polyp video dataset show that TSdetector achieves the highest polyp detection rate and outperforms other state-of-the-art methods. The code can be available at https://github.com/soleilssss/TSdetector.