Large Convolution Kernel Research Articles

Learning super-resolution and pyramidal convolution residual network for vehicle re-identification

Vehicle re-identification (Vehicle Re-ID) aims at retrieving and tracking the specified target vehicle with multiple other cameras, which can provide help in checking violations and catching fugitives, but there are still the following problems that need to be solved urgently. First, the existing collected Vehicle Re-ID data often have low resolution and blur in local regions, so that the Vehicle Re-ID algorithm cannot accurately extract subtle feature representations. In addition, small features are easy to cause the disappearance of features under the operation of a large convolution kernel, which makes the model unable to capture and learn subtle features, resulting in inaccurate judgment of vehicles. In this study, we propose a Vehicle Re-ID method based on super resolution and pyramidal convolution residual network. Firstly, a super-resolution image generation network leveraging generative adversarial networks (GANs) is proposed. This network employs both content loss and adversarial loss as optimization criteria, ensuring an efficient transformation from a low-resolution image into a super-resolution counterpart, while meticulously preserving intricate high-frequency details. Then, multi levels of pyramidal convolution operations are designed to generate multi-scale features, which can capture information on different scales. Moreover, the concept of residual learning is applied between the multi levels of pyramidal convolution operations to expedite model optimization and enhance recognition capabilities. Ultimately, the double pyramidal convolutions are meticulously employed on both the original image and the super-resolution image, yielding low-noise feature representations and intricate semantic information respectively. By seamlessly fusing these two diverse sources of information, the resultant combined features exhibit heightened discrimination capabilities and significantly bolster the robustness of image features. In order to verify the effectiveness of the proposed method, extensive experiments are carried out on VeRi-776 and VehicleID datasets. The experimental results show that the method proposed in this paper effectively captures the detail information of vehicle images, accurately distinguishes the subtle differences between different vehicles of the same type, and is superior to state-of-the-art methods.

MSMP-Net: A Multi-Scale Neural Network for End-to-End Monkeypox Virus Skin Lesion Classification

Monkeypox is a zoonotic disease caused by monkeypox virus infection. It is easily transmitted among people and poses a major threat to human health, making it of great significance in public health. Therefore, this paper proposes MSMP-Net, a multi-scale neural network for end-to-end monkeypox virus skin lesion classification ConvNeXt is used as the backbone network, and designs such as inverse bottleneck layers and large convolution kernels are used to enhance the network’s feature extraction capabilities. In order to effectively utilize the multi-level feature maps generated by the backbone network, a multi-scale feature fusion structure was designed. By fusing the deepest feature maps of multi-scale features, the model’s ability to represent monkeypox image features is enhanced. Experimental results show that the accuracy, precision, recall, and F1-score of this method on the MSLD v2.0 dataset are 87.03 ± 3.43%, 87.59 ± 3.37%, 87.03 ± 3.43%, and 86.58 ± 3.66%, respectively.

Real-Time ConvNext-Based U-Net with Feature Infusion for Egg Microcrack Detection

Real-time automatic detection of microcracks in eggs is crucial for ensuring egg quality and safety, yet rapid detection of micron-scale cracks remains challenging. This study introduces a real-time ConvNext-Based U-Net model with Feature Infusion (CBU-FI Net) for egg microcrack detection. Leveraging edge features and spatial continuity of cracks, we incorporate an edge feature infusion module in the encoder and design a multi-scale feature aggregation strategy in the decoder to enhance the extraction of both local details and global semantic information. By introducing large convolution kernels and depth-wise separable convolution from ConvNext, the model significantly reduces network parameters compared to the original U-Net. Additionally, a composite loss function is devised to address class imbalance issues. Experimental results on a dataset comprising over 3400 graded egg microcrack image patches demonstrate that CBU-FI Net achieves a reduction in parameters to one-third the amount in the original U-Net, with an inference speed of 21 ms per image (1 million pixels). The model achieves a Crack-IoU of 65.51% for microcracks smaller than 20 μm and a Crack-IoU and MIoU of 60.76% and 80.22%, respectively, for even smaller cracks (less than 5 μm), achieving high-precision, real-time detection of egg microcracks. Furthermore, on the publicly benchmarked CrackSeg9k dataset, CBU-FI Net achieves an inference speed of 4 ms for 400 × 400 resolution images, with an MIoU of 81.38%, proving the proposed method’s robustness and generalization capability across various cracks and complex backgrounds.

Power Quality Transient Disturbance Diagnosis Based on Dynamic Large Convolution Kernel and Multi-Level Feature Fusion Network

Power quality is an important metric for the normal operation of a power system, and the accurate identification of transient signals is of great significance for the improvement of power quality. The diverse types of power system transient signals and strong characteristic coupling brings new challenges to the analysis and identification of power system transient signals. In order to enhance the identification accuracy of transient signals, one method of power system transient signal identification is proposed based on a dynamic large convolution kernel and multilevel feature fusion network. First, the more fine-grained and more informative features of the transient signals are extracted by the dynamic large convolution kernel feature extraction module. Then, the multi-scale local features are adaptively fused by the multilevel feature fusion module. Finally, the fused features are reduced in dimension by the fully connected layer in the classification module and fed into the SoftMax layer for transient signal type detection. The proposed method can effectively improve the small receptive field problem of convolutional neural networks and the lack of ability of Transformer network in extracting local context information. Compared with five other power quality transient disturbance identification models, the experimental results show that the proposed method has better diagnostic accuracy and anti-noise capability.

THE IMPROVED ROAD DEFECT DETECTION ALGORITHM BASED ON YOLOv8s

Road defect detection is a key task to ensure road safety and repair damage in a timely manner. However, traditional manual inspection methods are inefficient and costly. To address such problems, a BL-YOLOv8 enhanced road defect detection algorithm is proposed. By integrating the BiFPN concept, the neck structure of the YOLOv8s model is reconstructed, reducing model parameters, computational load and overall size, and introducing a dynamic large convolution kernel attention mechanism (LSK-attention) to expand the model's receptive field and improve the accuracy of target detection. The experimental results show that on the road defect data set (KITTI), the number of parameters is reduced by 23.03 %, the amount of calculation is reduced by 5.85 %, and the average accuracy of mAP@0.5 is increased by 2.1 %, verifying the effectiveness of this method for automatic road defect detection technology effectiveness.

Image Super-resolution Networks based on Structure Reparameterized Convolution

The traditional image super resolution method based on convolutional neural network generally brings two problems: First, the model has only a single scale of receptive field, which cannot use the image information in a smaller range; Second, the size of the feature map will become smaller and smaller in the process of continuous convolution, and it is necessary to continuously carry out edge zeroing operation to maintain the original size, which leads to the loss of part of the edge information. To solve the above two problems, we propose a network structure based on structure re-parameterized convolution, which sets a small convolution kernel in parallel beside a large convolution kernel in the same layer, trains the two cores simultaneously, and finally merges the two cores. The experimental results show that in this way, we make the large convolution kernel could capture smaller information, which not only improves the high-frequency details of the reconstructed image but also avoids frequent edge filling to reduce the information density, and effectively speeds up the reasoning speed after the model is re-parameterized. Compared with the advanced methods, our method achieves good results.

TR-YOLO: A pig detection network based on YOLO V5n by combining self attention mechanism and large convolutional kernel

Under the highly valued environment of intelligent breeding, rapid and accurate detection of pigs in the breeding process can scientifically monitor the health of pigs and improve the welfare level of pigs. At present, the methods of live pig detection cannot complete the detection task in real time and accurately, so a pig detection model named TR-YOLO is proposed. Using cameras to collect data at the pig breeding site in Rongchang District, Chongqing City, LabelImg software is used to mark the position of pigs in the image, and data augmentation methods are used to expand the data samples, thus constructing a pig dataset. The lightweight YOLOv5n is selected as the baseline detection model. In order to complete the pig detection task more accurately, a C3DW module constructed by depth wise separable convolution with large convolution kernels is used to replace the C3 module in YOLOv5n, which enhances the receptive field of the whole detection model; a C3TR module constructed by Transformer structure is used to extract more refined global feature information. Contrast with the baseline model YOLOv5n, the new detection model does not increase additional computational load, and improves the accuracy of detection by 1.6 percentage points. Compared with other lightweight detection models, the new detection model has corresponding advantages in terms of parameter quantity, computational load, detection accuracy and so on. It can detect pigs in feeding more accurately while satisfying the real-time performance of target detection, providing an effective method for live monitoring and analysis of pigs at the production site.

Liquid-filled pipeline leak detection and localization based on multi-scale residual networks

Effective ways to improve the accuracy of liquid-filled pipeline leak detection are one of the key issues that need to be addressed urgently in a conservation-oriented society. Recently, pipeline leak detection methods based on deep learning have developed rapidly. To improve the learning ability of convolutional neural network for pipeline leak signal features and leak detection accuracy, a multi-scale residual networks (MSRNs) model is proposed in this paper for liquid-filled pipeline leak detection and localization. The model uses convolutional kernels of different scales to extract multiscale features of pipeline leakage signals based on deep residual networks (DRNs) and uses fully connected layers to fuse the features, thus improving the accuracy of pipeline leakage detection and localization. Among them, the large convolution kernel can acquire the low-frequency information of the signal due to its sizable perceptual field, the medium convolution kernel can capture the local and global features of the signal, and the small convolution kernel is more sensitive to the high-frequency information of the signal. Meanwhile, a pipeline leakage test platform is built to evaluate the proposed model. The test results show that the accuracy of leak detection and localization of MSRN model is 98.3%, which is better than that of single-scale DRN model. In addition, the proposed MSRN model is verified to have good generalization and noise immunity through testing and analyzing the leakage signals under different pressures and background noises.

Transformer with large convolution kernel decoder network for salient object detection in optical remote sensing images

Despite salient object detection in optical remote sensing images (ORSI-SOD) has made great strides in recent years, it is still a very challenging topic due to various scales and shapes of objects, cluttered backgrounds, and diverse imaging orientations. Most previous deep learning-based methods fails to effectively capture local and global features, resulting in ambiguous localization and semantic information and inaccurate detail and boundary prediction for ORSI-SOD. In this paper, we propose a novel Transformer with large convolutional kernel decoding network, named TLCKD-Net, which effectively models the long-range dependence that is indispensable for feature extraction of ORSI-SOD. First, we utilize Transformer backbone network to perceive global and local details of salient objects. Second, a large convolutional kernel decoding module based on self-attention mechanism is designed for different sizes of salient objects to extract feature information at different scales. Then, a large convolutional refinement and a Salient Feature Enhancement Module are used to recover and refine the saliency features to obtain high quality saliency maps. Extensive experiments on two public ORSI-SOD datasets show that our proposed method outperforms 16 state-of-the-art methods both qualitatively and quantitatively. In addition, a series of ablation studies demonstrate the effectiveness of different modules for ORSI-SOD. Our source code is publicly available at https://github.com/Dpw506/TLCKD-Net.

Micro-expression Recognition Based on DCBAM-EfficientNet Model

To address the problems of low accuracy of existing deep learning-based micro-expression recognition models, numerous network parameters, and the difficulty of mobile deployment of micro-expression recognition models, this paper proposes DCBAM-EfficientNet, a micro-expression recognition model that uses the lightweight network EfficientNet as the backbone network and incorporates the attention module. The network can guarantee the accuracy of micro-expression recognition with relatively few network parameters. The attention mechanism allows the more expressive micro-expression features to be highlighted, and the CBAM attention is improved into a DCBAM model, where the large convolution kernel in the spatial attention module of CBAM is replaced by a dilated convolution with the same receptive field, reducing the network parameters while better preserving the spatial features of the image. The integration of the DCBAM model into the main structure of EfficientNet enables better integration of contextual information. Data enhancement is used to process the micro-expression dataset to decrease the occurrence of overfitting and improve the generalization ability of the model. The results demonstrate that the optimized model DCBAM-EfficientNet can effectively promote the recognition accuracy of micro-expressions, significantly reduce the quantity and volume of model parameters, and provide a reference for the deployment of mobile micro-expression recognition models.

An Improved YOLOv5-Based Underwater Object-Detection Framework

To date, general-purpose object-detection methods have achieved a great deal. However, challenges such as degraded image quality, complex backgrounds, and the detection of marine organisms at different scales arise when identifying underwater organisms. To solve such problems and further improve the accuracy of relevant models, this study proposes a marine biological object-detection architecture based on an improved YOLOv5 framework. First, the backbone framework of Real-Time Models for object Detection (RTMDet) is introduced. The core module, Cross-Stage Partial Layer (CSPLayer), includes a large convolution kernel, which allows the detection network to precisely capture contextual information more comprehensively. Furthermore, a common convolution layer is added to the stem layer, to extract more valuable information from the images efficiently. Then, the BoT3 module with the multi-head self-attention (MHSA) mechanism is added into the neck module of YOLOv5, such that the detection network has a better effect in scenes with dense targets and the detection accuracy is further improved. The introduction of the BoT3 module represents a key innovation of this paper. Finally, union dataset augmentation (UDA) is performed on the training set using the Minimal Color Loss and Locally Adaptive Contrast Enhancement (MLLE) image augmentation method, and the result is used as the input to the improved YOLOv5 framework. Experiments on the underwater datasets URPC2019 and URPC2020 show that the proposed framework not only alleviates the interference of underwater image degradation, but also makes the mAP@0.5 reach 79.8% and 79.4% and improves the mAP@0.5 by 3.8% and 1.1%, respectively, when compared with the original YOLOv8 on URPC2019 and URPC2020, demonstrating that the proposed framework presents superior performance for the high-precision detection of marine organisms.

SDRC-YOLO: A Novel Foreign Object Intrusion Detection Algorithm in Railway Scenarios

Foreign object intrusion detection is vital to ensure the safety of railway transportation. Recently, object detection algorithms based on deep learning have been applied in a wide range of fields. However, in complex and volatile railway environments, high false detection, missed detection, and poor timeliness still exist in traditional object detection methods. To address these problems, an efficient railway foreign object intrusion detection approach SDRC-YOLO is proposed. First, a hybrid attention mechanism that fuses local representation ability is proposed to improve the identification accuracy of small targets. Second, DW-Decoupled Head is proposed to construct a mixed feature channel to improve localization and classification ability. Third, a large convolution kernel is applied to build a larger receptive field and improve the feature extraction capability of the network. In addition, the lightweight universal upsampling operator CARAFE is employed to sample the size and proportion of the intruding foreign body features in order to accelerate the convergence speed of the network. Experimental results show that, compared with the baseline YOLOv5s algorithm, SDRC-YOLO improved the mean average precision (mAP) by 2.8% and 1.8% on datasets RS and Pascal VOC 2012, respectively.

An Efficient and Lightweight Pigeon Age Detection Method Based on LN-STEP-YOLO

Deploying pigeon age detection model on edge equipment can solve the problem of video transmission delay and reduce the pressure of network transmission. Based on the deployment of edge devices, we made some improvements to You Only Look Once version 5 (YOLOv5) to form a new, lightweight, high-performance detector named LN-STEP-YOLO. In order to reduce the size of the model, we halved the number of channels in the YOLOv5s model. However, the decrease in the number of channels brings some problems, such as low global information acquisition, retention of image redundancy information, attention deficit, etc. To solve these problems, we did the following work. First, we proposed a new convolution structure with the effect of a large convolution kernel, StepConv. Second, a 2 × 2 convolution with step size 2 was used at the input to split each image into individual patches. Third, External Attention (EA) was introduced in the bottleneck structure. Fourth, a modified extremely separated convolutional block (XsepConv) was used for downsampling. Finally, we replaced the batch normalization (BN) of all non-downsampled layers with layer normalization (LN). The results showed that the improved algorithm outperformed generic lightweight networks such as Mixnet, Mobilenetv3, and Ghostnet to distinguish small-sized, overlapping pigeons, achieving 92.8% mean average precision (mAP) at about 17% of YOLOv5s parameters, 0.1% lower than that achieved by use of YOLOv5s. In addition, the improved method had 3.7G floating point operations (FLOPs) and 1.25G parameters, which allowed the detection of the growth stages of pigeons in real environments and provided a reference to guide placement of feeders in automated pigeon farming.

3D human pose estimation based on negative exponential reduction Gaussian kernel

Human pose estimation has become an important research direction in the field of motion recognition. 3D human pose estimation adds depth information to 2D pose estimation, which is more widely used. In this paper, the weight of each voxel is calculated in the 3D discrete space by projecting the joint point heatmap to directly estimate the 3D human pose. To improve the accuracy of 3D human pose estimation, the Gaussian kernel of heatmap with the variable size is reduced by a negative exponent in the process of training. The dilated convolution of a small convolution kernel is used to replace the large convolution kernel to solve the problem of large computation overhead when detecting key points in discrete 3D space. Experimental results show that this method is effective and can accurately estimate the 3D pose in multi view images.

A lightweight convolutional neural network model with receptive field block for C-shaped root canal detection in mandibular second molars

Rapid and accurate detection of a C-shaped root canal on mandibular second molars can assist dentists in diagnosis and treatment. Oral panoramic radiography is one of the most effective methods of determining the root canal of teeth. There are already some traditional methods based on deep learning to learn the characteristics of C-shaped root canal tooth images. However, previous studies have shown that the accuracy of detecting the C-shaped root canal still needs to be improved. And it is not suitable for implementing these network structures with limited hardware resources. In this paper, a new lightweight convolutional neural network is designed, which combined with receptive field block (RFB) for optimizing feature extraction. In order to optimize the hardware resource requirements of the model, a lightweight, multi-branch, convolutional neural network model was developed in this study. To improve the feature extraction ability of the model for C-shaped root canal tooth images, RFB has been merged with this model. RFB has achieved excellent results in target detection and classification. In the multiscale receptive field block, some small convolution kernels are used to replace the large convolution kernels, which allows the model to extract detailed features and reduce the computational complexity. Finally, the accuracy and area under receiver operating characteristics curve (AUC) values of C-shaped root canals on the image data of our mandibular second molars were 0.9838 and 0.996, respectively. The results show that the deep learning model proposed in this paper is more accurate and has lower computational complexity than many other similar studies. In addition, score-weighted class activation maps (Score-CAM) were generated to localize the internal structure that contributed to the predictions.

CAW: A Remote-Sensing Scene Classification Network Aided by Local Window Attention.

Remote-sensing image scene data contain a large number of scene images with different scales. Traditional scene classification algorithms based on convolutional neural networks are difficult to extract complex spatial distribution and texture information in images, resulting in poor classification results. In response to the above problems, we introduce the vision transformer network structure with strong global modeling ability into the remote-sensing image scene classification task. In this paper, the parallel network structure of the local-window self-attention mechanism and the equivalent large convolution kernel is used to realize the spatial-channel modeling of the network so that the network has better local and global feature extraction performance. Experiments on the RSSCN7 dataset and the WHU-RS19 dataset show that the proposed network can improve the accuracy of scene classification. At the same time, the effectiveness of the network structure in remote-sensing image classification tasks is verified through ablation experiments, confusion matrix, and heat map results comparison.

Bearing Running State Recognition Method Based on Feature-to-Noise Energy Ratio and Improved Deep Residual Shrinkage Network

Aiming at the difficulty in evaluating the degradation performance of rolling bearing and recognizing bearing running state, a new method based on feature-to-noise energy ratio (FNER) indicator and improved deep residual shrinkage network (IDRSN) model is proposed. First, Hilbert transform and fast Fourier transform are used to obtain the envelope spectrum of bearing run-to-failure test signal and the fault feature energy of the envelope spectrum amplitude is calculated. Subsequently, the total energy of the envelope signal is obtained according to the autocorrelation function (ACF). After that the ratio of the fault feature energy and the noise energy is used as the bearing performance degradation indicator. Moreover, bearing running states are divided according to FNER indicator and tagged bearing samples realized. Then, tagged samples are used to train the IDRSN, which introduces the densely connected network (DenseNet) to obtain the bearing running state recognition model. In order to improve the anti-interference ability, the DropBlock layer is introduced into the first large convolution kernel and the Dropout technique is introduced before the global average pooling layer. Finally, a failure test data is used to verify the feasibility of the FNER indicator and two bearing run-to-failure test data are used to verify the validity of the FNER indicator and the IDRSN recognition model. Through comparison and verification, the results show that the proposed FNER indicator is better than other indicators and IDRSN can more accurately recognize bearing running state recognition.

CodnNet: A lightweight CNN architecture for detection of COVID-19 infection

The application of Convolutional Neural Network (CNN) on the detection of COVID-19 infection has yielded favorable results. However, with excessive model parameters, the CNN detection of COVID-19 is low in recall, highly complex in computation. In this paper, a novel lightweight CNN model, CodnNet is proposed for quick detection of COVID-19 infection. CodnNet builds a more effective dense connections based on DenseNet network to make features highly reusable and enhances interactivity of local and global features. It also uses depthwise separable convolution with large convolution kernels instead of traditional convolution to improve the range of receptive field and enhances classification performance while reducing model complexity. The 5-Fold cross validation results on Kaggle’s COVID-19 Dataset showed that CodnNet has an average precision of 97.9%, recall of 97.4%, F1score of 97.7%, accuracy of 98.5%, mAP of 99.3%, and mAUC of 99.7%. Compared to the typical CNNs, CodnNet with fewer parameters and lower computational complexity has achieved better classification accuracy and generalization performance. Therefore, the CodnNet model provides a good reference for quick detection of COVID-19 infection.

Semantic Segmentation of Side-Scan Sonar Images with Few Samples

Underwater sensing and detection still rely heavily on acoustic equipment, known as sonar. As an imaging sonar, side-scan sonar can present a specific underwater situation in images, so the application scenario is comprehensive. However, the definition of side scan sonar is low; many objects are in the picture, and the scale is enormous. Therefore, the traditional image segmentation method is not practical. In addition, data acquisition is challenging, and the sample size is insufficient. To solve these problems, we design a semantic segmentation model of side-scan sonar images based on a convolutional neural network, which is used to realize the semantic segmentation of side-scan sonar images with few training samples. The model uses a large convolution kernel to extract large-scale features, adds a parallel channel using a small convolution kernel to obtain multi-scale features, and uses SE-block to focus on the weight of different channels. Finally, we verify the effect of the model on the self-collected side-scan sonar dataset. Experimental results show that, compared with the traditional lightweight semantic segmentation network, the model’s performance is improved, and the number of parameters is relatively small, which is easy to transplant to AUV.

Lightweight Smoke Recognition Based on Deep Convolution and Self-Attention

Deep convolutional networks have better smoke recognition performance. However, a lightweight network model and high recognition accuracy cannot be balanced when deployed on hardware with limited computing resources such as edge computing. Based on this background, we propose a novel smoke recognition network that combines convolutional networks (CNN) and self-attention. The core ideas of this framework are as follows: (1) Combine the depthwise convolution and asymmetric convolution of large convolution kernels to construct a lightweight CNN model, and realize multiscale extraction of feature information with slight model complexity. (2) Combined with the self-attention in transformer, a skip-connection branch is designed, which improves the feature extraction capability of the backbone network through parallel processing and fusion of feature map information. (3) Fusion multicomponent discrete cosine transform (DCT) is used to compress channel information and expand the ability of global average pooling (GAP) to aggregate feature maps. The proposed DCT-GAP improves the accuracy of the network without adding additional computational costs. Experimental results show that the proposed CSANet achieves an average accuracy of over 98.3% with 238 M FLOPs and 5.8 M parameters on the homemade smoke dataset, outperforming state-of-the-art competitors.