Convolution Module Research Articles

Defect detection of zipper tape based on lightweight deep learning network

Zipper tapes are essential in the garment industry, with defects significantly impacting the appearance of fabric products. Current methods for inspecting zipper tapes suffer from low detection accuracy. This paper proposes a lightweight, deep learning-based defect detection method using YOLOv5. The method includes a newly designed C3-CM module that integrates the Convolutional Block Attention Module (CBAM) into the backbone network, enhancing defect attention and feature extraction capabilities. Additionally, the lightweight GSConv convolution reduces parameter count while maintaining accuracy. The VoV-GSCSP module constructs deep-level feature fusion to boost the network’s expressive power. Using the Content-Aware Reassembly of Features (CARAFE), the method leverages underlying semantic information to predict and reorganize kernels, thereby enhancing upsampling effectiveness. Experimental results indicate that this method significantly improves accuracy compared to traditional approaches and outperforms current state-of-the-art algorithms in both accuracy and speed.

Multi-scale large kernel convolution and hybrid attention network for remote sensing image dehazing

Remote sensing (RS) image dehazing holds significant importance in enhancing the quality and information extraction capability of RS imagery. The enhancement in image dehazing quality has progressively advanced alongside the evolution of convolutional neural network (CNN). Due to the fixed receptive field of CNN, there is insufficient utilization of contextual information on haze features in multi-scale RS images. Additionally, the network fails to adequately extract both local and global information of haze features. In addressing the above problems, in this paper, we propose an RS image dehazing network based on multi-scale large kernel convolution and hybrid attention (MKHANet). The network is mainly composed of multi-scale large kernel convolution (MSLKC) module, hybrid attention (HA) module and feature fusion attention (FFA) module. The MSLKC module fully fuses the multi-scale information of features while enhancing the effective receptive field of the network by parallel multiple large kernel convolutions. To alleviate the problem of uneven distribution of haze and effectively extract the global and local information of haze features, the HA module is introduced by focusing on the importance of haze pixels at the channel level. The FFA module aims to boost the interaction of feature information between the network's deep and shallow layers. The subjective and objective experimental results on on multiple RS hazy image datasets illustrates that MKHANet surpasses existing state-of-the-art (SOTA) approaches. The source code is available at https://github.com/tohang98/MKHA_Net.

A radial basis deformable residual convolutional neural model embedded with local multi-modal feature knowledge and its application in cross-subject classification

In the spatial cognition and emotion recognition tasks based on electroencephalography (EEG), the signal information representation of the single modal is incomplete because of the significant inter-subject differences in the EEG signals, resulting in low generalization performance of classification models. To address this issue, we propose a radial basis deformable residual convolutional neural networks model embedded with local multi-modal feature knowledge (RBDRCNN-LMFK). The RBDRCNN leverages Euclidean distance alignment, deformable convolution, depth-separable convolution, and residual connection modules for effective EEG feature extraction. The embedded local feature knowledge algorithm enables the effective combination of multi-modal information. To further validate the effectiveness of the algorithm, we used the left-one-subject-out cross-validation algorithm on the virtual city walking (VCW), SJTU Emotion EEG Dataset IV (SEED IV), and Building block gesture recognition (BBGR) datasets for spatial cognition and emotion recognition tasks. The average accuracy of the VCW was 97.84 % using the kNN classifier, surpassing the Concate fusion method’s 96.62 %. The average accuracy for the SEED IV dataset was 87.56 %, higher than the Concate Fusion’s 69.97 %. On the BBGR dataset, the kNN classifier achieved an average accuracy of 87.80 %, compared to 85.31 % with the Concate fusion. The results show that the model enhances the recognition of biologically significant features in EEG signals by embedding local features and increases the correlation between different brain regions associated with the task. This work illustrates a promising direction of using deep learning models to discover effective task-related features from highly diverse EEG signals and enhance brain regions’ correlation through multi-modal knowledge embedding.

Cross-Modality Person Re-Identification Method with Joint-Modality Generation and Feature Enhancement.

In order to minimize the disparity between visible and infrared modalities and enhance pedestrian feature representation, a cross-modality person re-identification method is proposed, which integrates modality generation and feature enhancement. Specifically, a lightweight network is used for dimension reduction and augmentation of visible images, and intermediate modalities are generated to bridge the gap between visible images and infrared images. The Convolutional Block Attention Module is embedded into the ResNet50 backbone network to selectively emphasize key features sequentially from both channel and spatial dimensions. Additionally, the Gradient Centralization algorithm is introduced into the Stochastic Gradient Descent optimizer to accelerate convergence speed and improve generalization capability of the network model. Experimental results on SYSU-MM01 and RegDB datasets demonstrate that our improved network model achieves significant performance gains, with an increase in Rank-1 accuracy of 7.12% and 6.34%, as well as an improvement in mAP of 4.00% and 6.05%, respectively.

Research on traffic sign detection algorithm based on improved SSD in complex environments

Abstract In complex traffic sign environments, detection challenges include uneven size distribution, insufficient valid information, and difficulties in identifying targets under resource constraints, leading to missed and false detections. This study proposes an enhanced lightweight traffic sign detection algorithm based on SSD (Single Shot MultiBox Detector). By replacing the original backbone network with MobileNetV2, the model is streamlined to have fewer parameters, which improves generalization in complex environments. This modification significantly boosts the recall rate, achieving a better balance between precision and recall. By introducing the FPN(Feature Pyramid Network) combined with the CBAM(Convolutional Block Attention Module) attention mechanism, the detailed and semantic information between deep and shallow layers is fully integrated, reducing the loss of feature information, thus enhancing the strengthening of key information of traffic signs and the adaptability to different scales of traffic signs. Finally, by integrating the cross-attention mechanism, the algorithm's anti-interference ability in complex environments is improved, and the positioning accuracy of traffic signs is enhanced by capturing the dependency between different positions. Through ablation experiments and comparative experiments on a public traffic sign dataset, our improved SSD algorithm achieved an mAP of 89.97%. Compared with the original algorithm, the mAP increased by 12.41%, the recall rate increased by 18.38%, and the sum of precision and recall F1 increased by 14.6%. These improvements significantly enhance the performance of traffic sign detection in complex environments, thereby meeting the performance requirements of traffic sign detection.

Human Violence Detection in Videos Using Key Frame Identification and 3D CNN with Convolutional Block Attention Module

Human Violence Detection in Videos Using Key Frame Identification and 3D CNN with Convolutional Block Attention Module

Image Classification Based on Low-Level Feature Enhancement and Attention Mechanism

Deep learning-based image classification networks heavily rely on the extracted features. However, as the model becomes deeper, important features may be lost, resulting in decreased accuracy. To tackle this issue, this paper proposes an image classification method that enhances low-level features and incorporates an attention mechanism. The proposed method employs EfficientNet as the backbone network for feature extraction. Firstly, the Feature Enhancement Module quantifies and statistically processes low-level features from shallow layers, thereby enhancing the feature information. Secondly, the Convolutional Block Attention Module enhances the high-level features to improve the extraction of global features. Finally, the enhanced low-level features and global features are fused to supplement low-resolution global features with high-resolution details, further improving the model’s image classification ability. Experimental results illustrate that the proposed method achieves a Top-1 classification accuracy of 86.49% and a Top-5 classification accuracy of 96.90% on the ETH-Food101 dataset, 86.99% and 97.24% on the VireoFood-172 dataset, and 70.99% and 92.73% on the UEC-256 dataset. These results demonstrate that the proposed method outperforms existing methods in terms of classification performance.

Road underground defect detection in ground penetrating radar images based on an improved YOLOv5s model

Road underground defect detection plays a crucial role in assessing transportation infrastructure. Ground penetrating radar (GPR) serves as a widely used geophysical tool for this purpose. However, the traditional manual interpretation of GPR images heavily relies on the experience of the practitioner, leading to inefficiency and inaccuracies. To tackle these challenges, this paper proposes an automatic detection method for underground defects of roads based on an improved YOLOv5s model. First, the dense connection structure is integrated in the C3 module of the backbone to form the Dense-C3 module to enhance the capability of feature extraction. Subsequently, a convolutional block attention module (CBAM) is incorporated after each Dense-C3 module to refine features and enhance efficiency. Furthermore, the focal loss function is employed for the confidence loss to mitigate the impact of sample imbalance on detection performance. Experimental results demonstrate that the proposed model achieves a mean average precision (mAP) of 96.4% for synthetic data and 91.9% for real data, outperforming seven other models. The detection speed of the proposed model for real data reaches 51 frames per second, meeting the real-time detection requirements of road underground defects.

TP-Transfiner: high-quality segmentation network for tea pest.

Detecting and controlling tea pests promptly are crucial for safeguarding tea production quality. Due to the insufficient feature extraction ability of traditional CNN-based methods, they face challenges such as inaccuracy and inefficiency of detecting pests in dense and mimicry scenarios. This study proposes an end-to-end tea pest detection and segmentation framework, TeaPest-Transfiner (TP-Transfiner), based on Mask Transfiner to address the challenge of detecting and segmenting pests in mimicry and dense scenarios. In order to improve the feature extraction inability and weak accuracy of traditional convolution modules, this study proposes three strategies. Firstly, a deformable attention block is integrated into the model, which consists of deformable convolution and self-attention using the key content only term. Secondly, the FPN architecture in the backbone network is improved with a more effective feature-aligned pyramid network (FaPN). Lastly, focal loss is employed to balance positive and negative samples during the training period, and parameters are adapted to the dataset distribution. Furthermore, to address the lack of tea pest images, a dataset called TeaPestDataset is constructed, which contains 1,752 images and 29 species of tea pests. Experimental results on the TeaPestDataset show that the proposed TP-Transfiner model achieves state-of-the-art performance compared with other models, attaining a detection precision (AP50) of 87.211% and segmentation performance of 87.381%. Notably, the model shows a significant improvement in segmentation average precision (mAP) by 9.4% and a reduction in model size by 30% compared to the state-of-the-art CNN-based model Mask R-CNN. Simultaneously, TP-Transfiner's lightweight module fusion maintains fast inference speeds and a compact model size, demonstrating practical potential for pest control in tea gardens, especially in dense and mimicry scenarios.

The loose slipper fault diagnosis of variable-displacement pumps under time-varying operating conditions

Variable-displacement pumps (VDPs) are widely used as core power components of high-pressure hydraulic systems due to their superior power density. The loose slipper fault is a typical failure type of the VDP, which can cause premature damage or even unexpected shutdown. The on-site VDPs usually work under time-varying operating conditions (TVOCs), and their pressure varies with the changing speed. These unique characteristics pose a new challenge to the fault diagnosis (FD) of the VDPs. This study introduces an innovative FD method named the multi-scale attention mechanism residual network (MSARN). The MSARN is equipped with three well-chosen components: multi-scale convolution module (MSCM), attention mechanism (AM), and residual connection. The integration of these three components facilitates the efficient extraction and fusion of meaningful multi-scale fault features, thereby preventing the overfitting problem. The experimental results obtained on the VDP validate the effectiveness and accuracy of the MSARN. The suggested method trains the model using data collected under partial constant and representative operating conditions, enabling the achievement of FD under TVOCs.

Innovative lightweight deep learning architecture for enhanced rice pest identification

Pest detection is a crucial aspect of rice production. Accurate and timely identification of rice pests can assist farmers in taking prompt measures for control. To enhance the precision and real-time performance of rice pest detection, this paper introduces a novel YOLOv8-SCS architecture that integrates Space-to-Depth Convolution (SPD-Conv), Context Guided block (CG block), and Slide Loss. Initially, the original algorithm’s convolutional module is improved by introducing the SPD-Conv module, which reorganises the input channel dimensions into spatial dimensions, enabling the model to capture fine-grained pest features more efficiently while maintaining a lightweight model architecture. Subsequently, the CG block module is integrated into the CSPDarknet53 to 2-Stage FPN (C2f) structure, maintaining the models lightweight nature while enhancing its feature extraction capabilities. Finally, the Binary Cross-Entropy (BCE) is refined by incorporating the Slide Loss function, which encourages the model to focus more on challenging samples during training, thereby improving the model’s generalization across various samples. To validate the effectiveness of the improved algorithm, a series of experiments were conducted on a rice pest dataset. The results demonstrate that the proposed model outperforms the original YOLOv8 in rice pest detection, achieving an mAP of 87.9%, which is a 5.7% improvement over the original YOLOv8. The model also features a 44.1% reduction in parameter count and a decrease of 11.7 GFLOPs in computational requirements, meeting the demands for real-time detection.

DC-YOLOv5-based target detection algorithm for cervical vertebral maturation.

The cervical vertebral maturation (CVM) method is essential to determine the timing of orthodontic and orthopedic treatment. In this paper, a target detection model called DC-YOLOv5 is proposed to achieve fully automated detection and staging of CVM. A total of 1800 cephalometric radiographs were labeled and categorized based on the CVM stages. We introduced a model named DC-YOLOv5, optimized for the specific characteristics of CVM based on YOLOv5. This optimization includes replacing the original bounding box regression loss calculation method with Wise-IOU to address the issue of mutual interference between vertical and horizontal losses in Complete-IOU (CIOU), which made model convergence challenging. We incorporated the Res-dcn-head module structure to enhance the focus on small target features, improving the model's sensitivity to subtle sample differences. Additionally, we introduced the Convolutional Block Attention Module (CBAM) dual-channel attention mechanism to enhance focus and understanding of critical features, thereby enhancing the accuracy and efficiency of target detection. Loss functions, precision, recall, mean average precision (mAP), and F1 scores were used as the main algorithm evaluation metrics to assess the performance of these models. Furthermore, we attempted to analyze regions important for model predictions using gradient Class Activation Mapping (CAM) techniques. The final F1 scores of the DC-YOLOv5 model for CVM identification were 0.993, 0.994 for mAp0.5 and 0.943 for mAp0.5:0.95, with faster convergence, more accurate and more robust detection than the other four models. The DC-YOLOv5 algorithm shows high accuracy and robustness in CVM identification, which provides strong support for fast and accurate CVM identification and has a positive effect on the development of medical field and clinical diagnosis.

Ship Detection in Synthetic Aperture Radar Images Based on BiLevel Spatial Attention and Deep Poly Kernel Network

Synthetic aperture radar (SAR) is a technique widely used in the field of ship detection. However, due to the high ship density, fore-ground-background imbalance, and varying target sizes, achieving lightweight and high-precision multiscale ship object detection remains a significant challenge. In response to these challenges, this research presents YOLO-MSD, a multiscale SAR ship detection method. Firstly, we propose a Deep Poly Kernel Backbone Network (DPK-Net) that utilizes the Optimized Convolution (OC) Module to reduce data redundancy and the Poly Kernel (PK) Module to improve the feature extraction capability and scale adaptability. Secondly, we design a BiLevel Spatial Attention Module (BSAM), which consists of the BiLevel Routing Attention (BRA) and the Spatial Attention Module. The BRA is first utilized to capture global information. Then, the Spatial Attention Module is used to improve the network’s ability to localize the target and capture high-quality detailed information. Finally, we adopt a Powerful-IoU (P-IoU) loss function, which can adjust to the ship size adaptively, effectively guiding the anchor box to achieve faster and more accurate detection. Using HRSID and SSDD as experimental datasets, mAP of 90.2% and 98.8% are achieved, respectively, outperforming the baseline by 5.9% and 6.2% with a model size of 12.3 M. Furthermore, the network exhibits excellent performance across various ship scales.

Intelligent mechanical fault diagnosis using multiscale residual network and multisensor fusion

Abstract Mechanical faults in manufacturing systems need to be diagnosed accurately to ensure safety and cost savings. With the development of sensor technologies, data from multiple sensors is frequently utilized to assess the health of intricate industrial systems. In such cases, it is necessary to study the multisensor data based intelligent mechanical fault diagnosis method. First, the multisensor data is converted into grey images and then fused into a three-channel red-green-blue (RGB) image. Then, a multiscale with residual convolution module is proposed, which can extract multiscale deep features of the complex raw signal. Additionally, an attention module for channel and spatial attention is introduced to adaptively adjust the feature response values of each scale. Two datasets and a specific engineering application are used to validate the superiority of the network. The results show that the multisensor multiscale residual network outperforms other fault diagnosis networks in terms of fault identification accuracy, diagnostic efficiency, and applicability.

Design of a Leaf-Bottom Pest Control Robot with Adaptive Chassis and Adjustable Selective Nozzle

Pest control is an important guarantee for agricultural production. Pests are mostly light-avoiding and often gather on the bottom of crop leaves. However, spraying agricultural machinery mostly adopts top-down spraying, which suffers from low pesticide utilization and poor insect removal effect. Therefore, the upward spraying mode and intelligent nozzle have gradually become the research hotspot of precision agriculture. This paper designs a leaf-bottom pest control robot with adaptive chassis and adjustable selective nozzle. Firstly, the adaptive chassis is designed based on the MacPherson suspension, which uses shock absorption to drive the track to swing within a 30° angle. Secondly, a new type of cone angle adjustable selective nozzle was developed, which achieves adaptive selective precision spraying under visual guidance. Then, based on a convolutional block attention module (CBAM), the multi-CBAM-YOLOv5s network model was improved to achieve a 70% recognition rate of leaf-bottom spotted bad point in video streams. Finally, functional tests of the adaptive chassis and the adjustable selective spraying system were conducted. The data indicate that the adaptive chassis can adapt to diverse single-ridge requirements of soybeans and corn while protecting the ridge slopes. The selective spraying system achieves 70% precision in pesticide application, greatly reducing the use of pesticides. The scheme explores a ridge-friendly leaf-bottom pest control plan, providing a technical reference for improving spraying effect, reducing pesticide usage, and mitigating environmental pollution.

Repmono: a lightweight self-supervised monocular depth estimation architecture for high-speed inference

Self-supervised monocular depth estimation has always attracted attention because it does not require ground truth data. Designing a lightweight architecture capable of fast inference is crucial for deployment on mobile devices. The current network effectively integrates Convolutional Neural Networks (CNN) with Transformers, achieving significant improvements in accuracy. However, this advantage comes at the cost of an increase in model size and a significant reduction in inference speed. In this study, we propose a network named Repmono, which includes LCKT module with a large convolutional kernel and RepTM module based on the structural reparameterisation technique. With the combination of these two modules, our network achieves both local and global feature extraction with a smaller number of parameters and significantly enhances inference speed. Our network, with 2.31MB parameters, shows significant accuracy improvements over Monodepth2 in experiments on the KITTI dataset. With uniform input dimensions, our network’s inference speed is 53.7% faster than R-MSFM6, 60.1% faster than Monodepth2, and 81.1% faster than MonoVIT-small. Our code is available at https://github.com/txc320382/Repmono.

Multi-scale feature extraction and TrasMLP encoder module for ocean HABs segmentation

Due to tiny edge and texture details of harmful algae blooms(HABs), existing segmentation networks are not effective for HABs segmentation. In order to solve the above problems, this paper proposes a Multi-scale Feature extraction and TrasMLP Encoder Fusion-based network (MFTS). To tackle the complex morphological characteristics and the complex backgrounds of HABs, a TrasMlp module which can effectively identify long-range patterns and adapt network parameters is introduced, enabling accurately parsing of complex algae images. Secondly, the deep convolution module is constructed by combining deep separable convolution with a two-channel attention mechanism to separate the target region from the background. In addition, this paper proposes a Weighted Feature Fusion of Deep Convolution and INRS Encoder Module classification network(FDIR) is proposed to quantify the performance of the image segmentation network. The segmentation results on HABs dataset from AICO Lab show that our proposed MFTS model achieves a miou of 90.02%, outperforming the performance of classical segmentation networks such as U-Net and Mask R-CNN. Compared to original HABs dataset, the segmented result shows a 5.1% improvement in the classification accuracy of the FDIR model.

Attention improvement for data-driven analyzing fluorescence excitation-emission matrix spectra via interpretable attention mechanism

Analyzing three-dimensional excitation-emission matrix (3D-EEM) spectra through machine learning models has drawn increasing attention, whereas the reliability of these machine learning models remains unclear due to their “black box” nature. In this study, the convolutional neural network (CNN) for classifying numbers of fluorescent components in 3D-EEM spectra was interpreted by gradient-weighted class activation mapping (Grad-CAM), guided Grad-CAM, and structured attention graphs (SAGs). Results showed that the original CNN classifier with high classification accuracy may make a classification based on misleading attention to the non-fluorescence area in 3D-EEM spectra. By removing Rayleigh scatterings in 3D-EEM spectra and integrating convolutional block attention module (CBAM) in CNN classifiers, the correct attention of the trained CNN classifier with CBAM greatly increased from 17.6% to 57.2%. This work formulated strategies for improving CNN classifiers associated with environmental fields and would provide great help for water determination in both natural and artificial environments.

Adaptive loss-guided multi-stage residual ASPP for lesion segmentation and disease detection in cucumber under complex backgrounds

BackgroundIn complex agricultural environments, the presence of shadows, leaf debris, and uneven illumination can hinder the performance of leaf segmentation models for cucumber disease detection. This is further exacerbated by the imbalance in pixel ratios between background and lesion areas, which affects the accuracy of lesion extraction.ResultsAn original image segmentation framework, the LS-ASPP model, which utilizes a two-stage Atrous Spatial Pyramid Pooling (ASPP) approach combined with adaptive loss to address these challenges has been proposed. The Leaf-ASPP stage employs attention modules and residual structures to capture multi-scale semantic information and enhance edge perception, allowing for precise extraction of leaf contours from complex backgrounds. In the Spot-ASPP stage, we adjust the dilation rate of ASPP and introduce a Convolutional Attention Block Module (CABM) to accurately segment lesion areas.ConclusionsThe LS-ASPP model demonstrates improved performance in semantic segmentation accuracy under complex conditions, providing a robust solution for precise cucumber lesion segmentation. By focusing on challenging pixels and adapting to the specific requirements of agricultural image analysis, our framework has the potential to enhance disease detection accuracy and facilitate timely and effective crop management decisions.

FishFocusNet: An improved method based on YOLOv8 for underwater tropical fish identification

AbstractAccurately identifying tropical fish serves as a crucial indicator, offering an insight into the state of marine biodiversity and the condition of coral reef ecosystems. However, the current detection networks are prone to omission and misidentification due to occlusion between fish and the complex underwater environment. This paper proposes a modified approach named FishFocusNet, in which alterable kernel convolution modules, asymptotic feature pyramid network (AFPN), and Shape‐IoU are integrated into YOLOv8. To extract a more comprehensive set of fish features, AKConv modules with arbitrary kernel sizes are proposed to take the place of the conventional fixed‐shaped kernels in the backbone for downsampling. AFPN is adopted as the feature integration structure in the neck, which enhances feature fusion and adaptive spatial fusion between non‐adjacent layers. In the detector head, Shape‐IoU is employed to achieve precise localization of fish targets. The superiorities of these modifications are proved by ablation experiments and comparative experiments. The experimental results show that the optimized approach obtained an mAP of 81.8% accompanied by 2.4 MB parameters and 3.6 GB FLOPS. Meanwhile, compared with more complicated models of similar scale, the proposed method can enhance recognition accuracy to 84.2% and significantly reduce computational costs.