Yuanyuan Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yuanyuan Wang has a total of 61 co-authors. Here are the top co-authors by number of shared publications:
Haiyan Zhang
Huaiyin Institute of Technology, Huaian, China
10 shared publicationsShangbing Gao
College of Computer and Software Engineering, Huaiyin Institute of Technology, Huaian, China
6 shared publicationsHauwa Suleiman Abdullahi
School of Computer and Software Engineering, Huaiyin Institute of Technology, Huai’an, China
5 shared publicationsXiuchuan Chen
Affiliation College of Computer and Software Engineering, Huaiyin Institute of Technology, Huaian, China
5 shared publicationsTongtong Yin
Affiliation College of Computer and Software Engineering, Huaiyin Institute of Technology, Huaian, China
4 shared publicationsHu Zhao
Jiangsu Kesheng Xuanyi Technology Co., Ltd., Huaian, China
3 shared publications
Latest Publications
- Open Access
- 16 July 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (7) , 582
Image Super-Resolution Reconstruction Based on Hierarchical Feature Aggregation and Laplacian High-Frequency Compensation
Existing image super-resolution methods still suffer from limitations in edge-structure restoration, high-frequency texture preservation, and artifact suppression, which may lead to blurred contours and unnatural textures. To address these issues, this paper proposes an image super-resolution method based on hierarchical feature aggregation and Laplacian high-frequency compensation. First, a Hierarchical Feature Aggregation Attention Block (HFAB) is designed in the generator to progressively extract image features at different levels through multiple convolutional layers. A High-Frequency Variance Adaptive Channel Attention Block (HFVB) is further introduced to adaptively enhance key texture and edge information. Second, a Laplacian Adaptive Upsampling (LAU) module is developed to combine low-frequency content reconstruction with high-frequency detail compensation, thereby strengthening edge contours, preserving fine textures, and reducing artifacts. Finally, a Dissimilarity Structural Similarity Index Measure (DSSIM) loss is incorporated into the loss function to constrain local structural consistency and further improve the structural preservation and perceptual quality of reconstructed images. Experimental results on Set5, Set14, BSD100, and Urban100 show that, compared with SRGAN, the proposed method improves PSNR by 0.31 dB, 0.18 dB, 0.14 dB, and 0.14 dB, respectively, while reducing LPIPS by 0.0234, 0.0183, 0.0222, and 0.0222. These results indicate that the proposed method provides consistent improvements over SRGAN and achieves modest, metric-dependent gains over ESRGAN, suggesting an incremental enhancement in reconstruction accuracy and perceptual quality on both natural image benchmarks and complex urban scene datasets.
- 16 July 2026
- Published by MDPI AG in Algorithms
- Vol. 19 (7) , 582
Open AccessImage Super-Resolution Reconstruction Based on Hierarchical Feature Aggregation and Laplacian High-Frequency Compensation
Existing image super-resolution methods still suffer from limitations in edge-structure restoration, high-frequency texture preservation, and artifact suppression, which may lead to blurred contours and unnatural textures. To address these issues, this paper proposes an image super-resolution method based on hierarchical feature aggregation and Laplacian high-frequency compensation. First, a Hierarchical Feature Aggregation Attention Block (HFAB) is designed in the generator to progressively extract image features at different levels through multiple convolutional layers. A High-Frequency Variance Adaptive Channel Attention Block (HFVB) is further introduced to adaptively enhance key texture and edge information. Second, a Laplacian Adaptive Upsampling (LAU) module is developed to combine low-frequency content reconstruction with high-frequency detail compensation, thereby strengthening edge contours, preserving fine textures, and reducing artifacts. Finally, a Dissimilarity Structural Similarity Index Measure (DSSIM) loss is incorporated into the loss function to constrain local structural consistency and further improve the structural preservation and perceptual quality of reconstructed images. Experimental results on Set5, Set14, BSD100, and Urban100 show that, compared with SRGAN, the proposed method improves PSNR by 0.31 dB, 0.18 dB, 0.14 dB, and 0.14 dB, respectively, while reducing LPIPS by 0.0234, 0.0183, 0.0222, and 0.0222. These results indicate that the proposed method provides consistent improvements over SRGAN and achieves modest, metric-dependent gains over ESRGAN, suggesting an incremental enhancement in reconstruction accuracy and perceptual quality on both natural image benchmarks and complex urban scene datasets.
- Open Access
- 2 December 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Multiscale scene parsing network
To address the core challenge faced by existing lightweight scene parsing networks—balancing multiscale feature representation precision and computational efficiency (rather than “difficulties in extracting multi-scale information”)—this paper proposes MSPNet, a lightweight multiscale scene parsing network. The network adopts StarNet as the backbone to leverage its efficient low-to-high dimensional feature transformation capability, and innovatively embeds the Efficient Pixel Localization Attention (EPLA) module into the PSPNet architecture. Unlike simple module stacking, the EPLA module integrates two synergistic submodules: ELA (Efficient Localization Attention) and PagFM (Pyramid Attention-Guided Feature Module). The ELA module uses a dynamic weight allocation mechanism to achieve precise pixel-level feature localization while reducing attention computation overhead by 38%; the PagFM module constructs a hierarchical pyramid fusion architecture, adaptively guiding cross-scale feature integration to enhance small-target representation. Additionally, MSPNet incorporates depthwise separable convolutions and channel reparameterization techniques, further optimizing model compactness. Experimental results on the Pascal VOC2012 validation set show that MSPNet achieves a mean Intersection over Union (mIoU) of 87.19%, a 1.79% improvement over PSPNet. With GFLOPs (9.7G for StarNet-s4 backbone) and parameter counts (7.4 M) comparable to the MobileNet series, MSPNet outperforms contemporary lightweight SOTA models in both accuracy and efficiency, providing an effective solution for real-time semantic segmentation on resource-constrained mobile devices. The code for MSPNet is available at https://github.com/Eric-863/MSPnet .
- 2 December 2025
- Published by Springer Nature in Scientific Reports
- Vol. 16 (1)
Open AccessMultiscale scene parsing network
To address the core challenge faced by existing lightweight scene parsing networks—balancing multiscale feature representation precision and computational efficiency (rather than “difficulties in extracting multi-scale information”)—this paper proposes MSPNet, a lightweight multiscale scene parsing network. The network adopts StarNet as the backbone to leverage its efficient low-to-high dimensional feature transformation capability, and innovatively embeds the Efficient Pixel Localization Attention (EPLA) module into the PSPNet architecture. Unlike simple module stacking, the EPLA module integrates two synergistic submodules: ELA (Efficient Localization Attention) and PagFM (Pyramid Attention-Guided Feature Module). The ELA module uses a dynamic weight allocation mechanism to achieve precise pixel-level feature localization while reducing attention computation overhead by 38%; the PagFM module constructs a hierarchical pyramid fusion architecture, adaptively guiding cross-scale feature integration to enhance small-target representation. Additionally, MSPNet incorporates depthwise separable convolutions and channel reparameterization techniques, further optimizing model compactness. Experimental results on the Pascal VOC2012 validation set show that MSPNet achieves a mean Intersection over Union (mIoU) of 87.19%, a 1.79% improvement over PSPNet. With GFLOPs (9.7G for StarNet-s4 backbone) and parameter counts (7.4 M) comparable to the MobileNet series, MSPNet outperforms contemporary lightweight SOTA models in both accuracy and efficiency, providing an effective solution for real-time semantic segmentation on resource-constrained mobile devices. The code for MSPNet is available at https://github.com/Eric-863/MSPnet .
- Open Access
- 29 August 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (8) , e0330039
SCP-DETR: A efficient small-object-enhanced feature pyramid approach for PCB defect detection
Defects generated during PCB manufacturing, transportation, and storage seriously impact the quality and performance of electronic components. However, detection accuracy is limited due to excessive background interference and the small size of defect targets. To alleviate these issues, this paper proposes an improved PCB defect detection method based on RT-DETR, named SCP-DETR. Firstly, to effectively detect small targets, the S2 feature layer is incorporated into the neck feature fusion. While this improves detection capability, it also introduces considerable computational overhead. To mitigate this, we use SPDConv (Space-to-Depth Convolution) to process the S2 feature layer, reducing the computational complexity. The processed S2 feature layer is then fused with the S3 feature layer and higher-level features. Subsequently, we feed these features into a specially designed CO-Fusion module. By embedding our proposed CSPOKM(CSP Omni-Kernel Module) into the original fusion module, the CO-Fusion module effectively learns feature representations from global to local scales, ultimately enhancing small-target detection performance. Finally, downsampling operations are replaced with PSConv(Pinwheel-shaped Convolution), which better accommodates the Gaussian spatial pixel distributions of subtle small targets. Experimental results demonstrate that the proposed method achieves an mAP@0.5 of 97%, surpassing RT-DETR-r18 by 3%, and an mAP@0.5:0.95 of 53.4%, representing an improvement of 2.2%. Additionally, compared with the recently released YOLOv11m, our method improves mAP@0.5 by 5.6%. These results demonstrate the superior performance of the proposed method in PCB defect detection, which holds significant implications for industrial production. The code is available at https://github.com/Yttong-rr/SCPDETR/tree/master.
- 29 August 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (8) , e0330039
Open AccessSCP-DETR: A efficient small-object-enhanced feature pyramid approach for PCB defect detection
Defects generated during PCB manufacturing, transportation, and storage seriously impact the quality and performance of electronic components. However, detection accuracy is limited due to excessive background interference and the small size of defect targets. To alleviate these issues, this paper proposes an improved PCB defect detection method based on RT-DETR, named SCP-DETR. Firstly, to effectively detect small targets, the S2 feature layer is incorporated into the neck feature fusion. While this improves detection capability, it also introduces considerable computational overhead. To mitigate this, we use SPDConv (Space-to-Depth Convolution) to process the S2 feature layer, reducing the computational complexity. The processed S2 feature layer is then fused with the S3 feature layer and higher-level features. Subsequently, we feed these features into a specially designed CO-Fusion module. By embedding our proposed CSPOKM(CSP Omni-Kernel Module) into the original fusion module, the CO-Fusion module effectively learns feature representations from global to local scales, ultimately enhancing small-target detection performance. Finally, downsampling operations are replaced with PSConv(Pinwheel-shaped Convolution), which better accommodates the Gaussian spatial pixel distributions of subtle small targets. Experimental results demonstrate that the proposed method achieves an mAP@0.5 of 97%, surpassing RT-DETR-r18 by 3%, and an mAP@0.5:0.95 of 53.4%, representing an improvement of 2.2%. Additionally, compared with the recently released YOLOv11m, our method improves mAP@0.5 by 5.6%. These results demonstrate the superior performance of the proposed method in PCB defect detection, which holds significant implications for industrial production. The code is available at https://github.com/Yttong-rr/SCPDETR/tree/master.
- Open Access
- 30 May 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-21
Research on a traffic flow statistical algorithm based on YBOVDT and SAM2
In the process of urbanization, traffic flow statistics are of great significance to traffic management. Existing traffic flow statistics solutions suffer from incomplete functionality and lack effective solutions for core issues. The closed-set object detection algorithms they employ can only perform detections based on fixed categories, which leads to limited recognition scope and weak model generalization ability.Moreover, the tracking algorithms used are unstable and have low computational efficiency. To address these challenges, this paper proposes a traffic flow statistical method based on YBOVDT(YOLO-World and BOT-SORT-Open Vocabulary Detection and Tracking)and SAM2.Specifically, in the method, this paper proposes a “Traffic Flow Data Processing and Analysis” module, aiming to optimize and supplement the five core functions required for traffic flow statistics tasks, thereby making the functions of the entire solution more comprehensive.In addition, this paper combines the latest open set object detection and tracking algorithms to enhance the recognition ability and tracking stability of traffic objects. In this study, a custom dataset was used to train existing traffic flow statistics models.The experimental results showed that the YOLO-World model achieved a precision of 76.99% and an mAP50 of 70.08%. A comparative analysis with YOLO-v3,YOLO-v5, YOLO-v6,and YOLO-v8 algorithms indicated that, while balancing spatial and temporal resource consumption and accuracy, the proposed algorithm offers higher recognition accuracy and environmental adaptability. The experimental results further validated that this method demonstrates significant improvements in handling traffic flow statistics tasks in complex traffic environments.
- 30 May 2025
- Published by Springer Nature in Scientific Reports
- Vol. 15 (1) , 1-21
Open AccessResearch on a traffic flow statistical algorithm based on YBOVDT and SAM2
In the process of urbanization, traffic flow statistics are of great significance to traffic management. Existing traffic flow statistics solutions suffer from incomplete functionality and lack effective solutions for core issues. The closed-set object detection algorithms they employ can only perform detections based on fixed categories, which leads to limited recognition scope and weak model generalization ability.Moreover, the tracking algorithms used are unstable and have low computational efficiency. To address these challenges, this paper proposes a traffic flow statistical method based on YBOVDT(YOLO-World and BOT-SORT-Open Vocabulary Detection and Tracking)and SAM2.Specifically, in the method, this paper proposes a “Traffic Flow Data Processing and Analysis” module, aiming to optimize and supplement the five core functions required for traffic flow statistics tasks, thereby making the functions of the entire solution more comprehensive.In addition, this paper combines the latest open set object detection and tracking algorithms to enhance the recognition ability and tracking stability of traffic objects. In this study, a custom dataset was used to train existing traffic flow statistics models.The experimental results showed that the YOLO-World model achieved a precision of 76.99% and an mAP50 of 70.08%. A comparative analysis with YOLO-v3,YOLO-v5, YOLO-v6,and YOLO-v8 algorithms indicated that, while balancing spatial and temporal resource consumption and accuracy, the proposed algorithm offers higher recognition accuracy and environmental adaptability. The experimental results further validated that this method demonstrates significant improvements in handling traffic flow statistics tasks in complex traffic environments.
- Open Access
- 25 April 2025
- Published by Springer Nature in Journal of Big Data
- Vol. 12 (1) , 1-22
FEL-FRN: fusion ECA long-CLIP feature reconstruction network for few-shot classification
This paper focuses on the problem of the feature reconstruction network (FRN) facing difficulty in reconstructing the query set in fine-grained classification tasks when the objects in the support set have multiple attributes. To address this problem, we propose a model called FEL-FRN (fusion ECA Long-CLIP feature reconstruction network). First, we use FRN to deeply reconstruct feature maps, replacing the traditional method of using cosine similarity for category average aggregation. Moreover, we introduce the efficient channel attention (ECA) attention mechanism into the FRN to improve the model’s ability to extract key features. Second, by introducing Long-CLIP to assist FRN recognition, the Long-CLIP model with a wide range of image recognition and understanding capabilities is obtained. The model does not require any task-specific fine-tuning data and can be combined with category text prediction. Finally, in each training task, the prediction results of different branches are fused. The Long-CLIP model can effectively compensate for the problem of poor prediction caused by large differences between the reconstructed support images and the poor quality of the reconstructed images, whereas the FRN reconstruction network compensates for the lack of precision in Long-CLIP direct prediction through reconstructed predictions, achieving complementary advantages. The experimental results show that this FEL-FRN method not only achieves good results on CUB-200–2011 and Oxford 102 flowers but also uses 5way5shot as a support set on the car dataset Stanford_Cars and the aircraft dataset FGVC_Aircraft, which have large attribute differences, with accuracies of 96.025% and 81.479%, respectively. The results show that the performance is improved compared with that of the FRN model strategy used alone. https://github.com/feiyeha/FEL-FRN
- 25 April 2025
- Published by Springer Nature in Journal of Big Data
- Vol. 12 (1) , 1-22
Open AccessFEL-FRN: fusion ECA long-CLIP feature reconstruction network for few-shot classification
This paper focuses on the problem of the feature reconstruction network (FRN) facing difficulty in reconstructing the query set in fine-grained classification tasks when the objects in the support set have multiple attributes. To address this problem, we propose a model called FEL-FRN (fusion ECA Long-CLIP feature reconstruction network). First, we use FRN to deeply reconstruct feature maps, replacing the traditional method of using cosine similarity for category average aggregation. Moreover, we introduce the efficient channel attention (ECA) attention mechanism into the FRN to improve the model’s ability to extract key features. Second, by introducing Long-CLIP to assist FRN recognition, the Long-CLIP model with a wide range of image recognition and understanding capabilities is obtained. The model does not require any task-specific fine-tuning data and can be combined with category text prediction. Finally, in each training task, the prediction results of different branches are fused. The Long-CLIP model can effectively compensate for the problem of poor prediction caused by large differences between the reconstructed support images and the poor quality of the reconstructed images, whereas the FRN reconstruction network compensates for the lack of precision in Long-CLIP direct prediction through reconstructed predictions, achieving complementary advantages. The experimental results show that this FEL-FRN method not only achieves good results on CUB-200–2011 and Oxford 102 flowers but also uses 5way5shot as a support set on the car dataset Stanford_Cars and the aircraft dataset FGVC_Aircraft, which have large attribute differences, with accuracies of 96.025% and 81.479%, respectively. The results show that the performance is improved compared with that of the FRN model strategy used alone. https://github.com/feiyeha/FEL-FRN
- Open Access
- 7 April 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (4) , e0318033
Lightweight PCB defect detection method based on SCF-YOLO
Addressing the issues of large model size and slow detection speed in real-time defect detection in complex scenarios of printed circuit boards (PCBs), this study proposes a new lightweight defect detection model called SCF-YOLO. The aim of SCF-YOLO is to solve the problem of resource limitation in algorithm deployment. SCF-YOLO utilizes the more compact and lightweight MobileNet as the feature extraction network, which effectively reduces the number of model parameters and significantly improves the inference speed. Additionally, the model introduces a learnable weighted feature fusion module in the neck, which enhances the expression of features at multiple scales and different levels, thus improving the focus on key features. Furthermore, a novel SCF module (Synthesis C2f) is proposed to enhance the model’s ability to capture high-level semantic features. During the training process, a combined loss function that combines CIoU and GIoU is used to effectively balance the optimization of different objectives and ensure the precise location of defects. Experimental results demonstrate that compared to the YOLOv8 algorithm, SCF-YOLO reduces the number of parameters by 25% and improves the detection speed by up to 60%. This provides a fast, accurate, and efficient solution for defect detection of PCBs in industrial production.
- 7 April 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (4) , e0318033
Open AccessLightweight PCB defect detection method based on SCF-YOLO
Addressing the issues of large model size and slow detection speed in real-time defect detection in complex scenarios of printed circuit boards (PCBs), this study proposes a new lightweight defect detection model called SCF-YOLO. The aim of SCF-YOLO is to solve the problem of resource limitation in algorithm deployment. SCF-YOLO utilizes the more compact and lightweight MobileNet as the feature extraction network, which effectively reduces the number of model parameters and significantly improves the inference speed. Additionally, the model introduces a learnable weighted feature fusion module in the neck, which enhances the expression of features at multiple scales and different levels, thus improving the focus on key features. Furthermore, a novel SCF module (Synthesis C2f) is proposed to enhance the model’s ability to capture high-level semantic features. During the training process, a combined loss function that combines CIoU and GIoU is used to effectively balance the optimization of different objectives and ensure the precise location of defects. Experimental results demonstrate that compared to the YOLOv8 algorithm, SCF-YOLO reduces the number of parameters by 25% and improves the detection speed by up to 60%. This provides a fast, accurate, and efficient solution for defect detection of PCBs in industrial production.
- Open Access
- 4 April 2025
- Published by Springer Nature in Research Square
MSPNet: Multiscale Scene Parsing Network
To address the challenges faced by existing scene parsing networks, such as high parameter counts, high computational complexity, and difficulties in achieving effective multiscalefeature representation, this paper proposes a lightweight multiscalescene parsing network called MSPNet. By adopting StarNet as the backbone network and embedding the efficient pixel localization attention (EPLA) module into PSPNet, the proposed network achieves a significant performance improvement. The EPLA module integrates two submodules: ELA and PagFM. The ELA module usesa dynamic weight allocation mechanism to achieve precise pixel-level feature localization, effectively reducing the computational overhead of attention mechanisms. Moreover, the PagFM module constructs a hierarchical pyramid feature fusion architecture, guiding and fusing features at different scales. Through their synergistic interaction, these two modules greatly enhance the network's ability to represent multiscale targets. Additionally, MSPNet leverages depthwise separable convolutions and channel reparameterization techniques to maintain a lightweight design while ensuring computational efficiency. The experimental results demonstrate that MSPNet performs exceptionally well on the Pascal VOC2012 validation set, achieving a 1.79% improvement in the mean intersection over union (mIoU) overPSPNet. Through the collaborative optimization of innovative modules, MSPNet excels in lightweight design, with GFLOPS and parameter counts comparable to those of the MobileNet series, providing an efficient solution for real-time semantic segmentation on mobile devices.
- 4 April 2025
- Published by Springer Nature in Research Square
Open AccessMSPNet: Multiscale Scene Parsing Network
To address the challenges faced by existing scene parsing networks, such as high parameter counts, high computational complexity, and difficulties in achieving effective multiscalefeature representation, this paper proposes a lightweight multiscalescene parsing network called MSPNet. By adopting StarNet as the backbone network and embedding the efficient pixel localization attention (EPLA) module into PSPNet, the proposed network achieves a significant performance improvement. The EPLA module integrates two submodules: ELA and PagFM. The ELA module usesa dynamic weight allocation mechanism to achieve precise pixel-level feature localization, effectively reducing the computational overhead of attention mechanisms. Moreover, the PagFM module constructs a hierarchical pyramid feature fusion architecture, guiding and fusing features at different scales. Through their synergistic interaction, these two modules greatly enhance the network's ability to represent multiscale targets. Additionally, MSPNet leverages depthwise separable convolutions and channel reparameterization techniques to maintain a lightweight design while ensuring computational efficiency. The experimental results demonstrate that MSPNet performs exceptionally well on the Pascal VOC2012 validation set, achieving a 1.79% improvement in the mean intersection over union (mIoU) overPSPNet. Through the collaborative optimization of innovative modules, MSPNet excels in lightweight design, with GFLOPS and parameter counts comparable to those of the MobileNet series, providing an efficient solution for real-time semantic segmentation on mobile devices.
- Open Access
- 1 January 2025
- Published by Tech Science Press in Computers, Materials & Continua
- Vol. 86 (3) , 1-10
Steel Surface Defect Detection via the Multiscale Edge Enhancement Method
To solve the false detection and missed detection problems caused by various types and sizes of defects in the detection of steel surface defects, similar defects and background features, and similarities between different defects, this paper proposes a lightweight detection model named multiscale edge and squeeze-and-excitation attention detection network (MSESE), which is built upon the You Only Look Once version 11 nano (YOLOv11n). To address the difficulty of locating defect edges, we first propose an edge enhancement module (EEM), apply it to the process of multiscale feature extraction, and then propose a multiscale edge enhancement module (MSEEM). By obtaining defect features from different scales and enhancing their edge contours, the module uses the dual-domain selection mechanism to effectively focus on the important areas in the image to ensure that the feature images have richer information and clearer contour features. By fusing the squeeze-and-excitation attention mechanism with the EEM, we obtain a lighter module that can enhance the representation of edge features, which is named the edge enhancement module with squeeze-and-excitation attention (EEMSE). This module was subsequently integrated into the detection head. The enhanced detection head achieves improved edge feature enhancement with reduced computational overhead, while effectively adjusting channel-wise importance and further refining feature representation. Experiments on the NEU-DET dataset show that, compared with the original YOLOv11n, the improved model achieves improvements of 4.1% and 2.2% in terms of mAP@0.5 and mAP@0.5:0.95, respectively, and the GFLOPs value decreases from the original value of 6.4 to 6.2. Furthermore, when compared to current mainstream models, Mamba-YOLOT and RTDETR-R34, our method achieves superior performance with 6.5% and 8.9% higher mAP@0.5, respectively, while maintaining a more compact parameter footprint. These results collectively validate the effectiveness and efficiency of our proposed approach.
- 1 January 2025
- Published by Tech Science Press in Computers, Materials & Continua
- Vol. 86 (3) , 1-10
Open AccessSteel Surface Defect Detection via the Multiscale Edge Enhancement Method
To solve the false detection and missed detection problems caused by various types and sizes of defects in the detection of steel surface defects, similar defects and background features, and similarities between different defects, this paper proposes a lightweight detection model named multiscale edge and squeeze-and-excitation attention detection network (MSESE), which is built upon the You Only Look Once version 11 nano (YOLOv11n). To address the difficulty of locating defect edges, we first propose an edge enhancement module (EEM), apply it to the process of multiscale feature extraction, and then propose a multiscale edge enhancement module (MSEEM). By obtaining defect features from different scales and enhancing their edge contours, the module uses the dual-domain selection mechanism to effectively focus on the important areas in the image to ensure that the feature images have richer information and clearer contour features. By fusing the squeeze-and-excitation attention mechanism with the EEM, we obtain a lighter module that can enhance the representation of edge features, which is named the edge enhancement module with squeeze-and-excitation attention (EEMSE). This module was subsequently integrated into the detection head. The enhanced detection head achieves improved edge feature enhancement with reduced computational overhead, while effectively adjusting channel-wise importance and further refining feature representation. Experiments on the NEU-DET dataset show that, compared with the original YOLOv11n, the improved model achieves improvements of 4.1% and 2.2% in terms of mAP@0.5 and mAP@0.5:0.95, respectively, and the GFLOPs value decreases from the original value of 6.4 to 6.2. Furthermore, when compared to current mainstream models, Mamba-YOLOT and RTDETR-R34, our method achieves superior performance with 6.5% and 8.9% higher mAP@0.5, respectively, while maintaining a more compact parameter footprint. These results collectively validate the effectiveness and efficiency of our proposed approach.
- Open Access
- 12 November 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-21
A steel defect detection method based on edge feature extraction via the Sobel operator
Scratches and cracks in steel severely affect its service life and performance. However, owing to the irregular shapes and sizes of steel surface defects, defects within the same class may be different, whereas defects between classes may be similar. Existing methods focus only on spatial information, resulting in low detection accuracy. To alleviate these problems, this paper proposes the ECDY (EIFEM CARAFE DyHead) network to enhance the detection capability of steel defects. We first design a feature extraction module that focuses on the edge information of feature contours. This module uses the Sobel operator to extract the edge information of a feature and fuses it with the overall spatial information so that richer semantic information can be obtained. The module has improved accuracy in the YOLOv5, YOLOv8, and YOLOv10 versions, and uses fewer parameters and calculations. In particular, in YOLOv8x, mAP@0.5 increased by 2.5%, and the number of parameters was reduced by 12.4 M. Second, to retain the detailed information in the feature pyramid, and to better reconstruct features, we choose the content-aware reassembly feature method (CARAFE) as the upsampling method. Finally, the detection head was replaced with a dynamic unified detection head (DyHead) to adapt to different defect sizes and different task requirements. Compared with YOLOv8s, the proposed method improves precision by 1.6%, recall by 4%, and mAP@0.5 by 4%. This value is 4.2% higher than the mAP@0.5 of the current SOTA model RT-DETR-L in the field of object detection and has 23.2 M fewer parameters.
- 12 November 2024
- Published by Springer Nature in Scientific Reports
- Vol. 14 (1) , 1-21
Open AccessA steel defect detection method based on edge feature extraction via the Sobel operator
Scratches and cracks in steel severely affect its service life and performance. However, owing to the irregular shapes and sizes of steel surface defects, defects within the same class may be different, whereas defects between classes may be similar. Existing methods focus only on spatial information, resulting in low detection accuracy. To alleviate these problems, this paper proposes the ECDY (EIFEM CARAFE DyHead) network to enhance the detection capability of steel defects. We first design a feature extraction module that focuses on the edge information of feature contours. This module uses the Sobel operator to extract the edge information of a feature and fuses it with the overall spatial information so that richer semantic information can be obtained. The module has improved accuracy in the YOLOv5, YOLOv8, and YOLOv10 versions, and uses fewer parameters and calculations. In particular, in YOLOv8x, mAP@0.5 increased by 2.5%, and the number of parameters was reduced by 12.4 M. Second, to retain the detailed information in the feature pyramid, and to better reconstruct features, we choose the content-aware reassembly feature method (CARAFE) as the upsampling method. Finally, the detection head was replaced with a dynamic unified detection head (DyHead) to adapt to different defect sizes and different task requirements. Compared with YOLOv8s, the proposed method improves precision by 1.6%, recall by 4%, and mAP@0.5 by 4%. This value is 4.2% higher than the mAP@0.5 of the current SOTA model RT-DETR-L in the field of object detection and has 23.2 M fewer parameters.
- Open Access
- 30 August 2024
- Published by Frontiers Media SA in Frontiers in Physics
- Vol. 12, 1451165
Multiclass small target detection algorithm for surface defects of chemicals special steel
Introduction: Chemical special steels are widely used in chemical equipment manufacturing and other fields, and small defects on its surface (such as cracks and punches) are easy to cause serious accidents in harsh environments.Methods: In order to solve this problem, this paper proposes an improved defect detection algorithm for chemical special steel based on YOLOv8. Firstly, in order to effectively capture local and global information, a ParC2Net (Parallel-C2f) structure is proposed for feature extraction, which can accurately capture the subtle features of steel defects. Secondly, the loss function is adjusted to MPD-IOU, and its dynamic non-monotonic focusing characteristics are used to effectively solve the overfitting problem of the bounding box of low-quality targets. In addition, RepGFPN is used to fuse multi-scale features, deepen the interaction between semantics and spatial information, and significantly improve the efficiency of cross-layer information transmission. Finally, the RexSE-Head (ResNeXt-Squeeze-Excitation) design is adopted to enhance the positioning accuracy of small defect targets.Results and discussion: The experimental results show that the mAP@0.5 of the improved model reaches 93.5%, and the number of parameters is only 3.29M, which realizes the high precision and high response performance of the detection of small defects in chemical special steels, and highlights the practical application value of the model. The code is available at https://github.com/improvment/prs-yolo.
- 30 August 2024
- Published by Frontiers Media SA in Frontiers in Physics
- Vol. 12, 1451165
Open AccessMulticlass small target detection algorithm for surface defects of chemicals special steel
Introduction: Chemical special steels are widely used in chemical equipment manufacturing and other fields, and small defects on its surface (such as cracks and punches) are easy to cause serious accidents in harsh environments.Methods: In order to solve this problem, this paper proposes an improved defect detection algorithm for chemical special steel based on YOLOv8. Firstly, in order to effectively capture local and global information, a ParC2Net (Parallel-C2f) structure is proposed for feature extraction, which can accurately capture the subtle features of steel defects. Secondly, the loss function is adjusted to MPD-IOU, and its dynamic non-monotonic focusing characteristics are used to effectively solve the overfitting problem of the bounding box of low-quality targets. In addition, RepGFPN is used to fuse multi-scale features, deepen the interaction between semantics and spatial information, and significantly improve the efficiency of cross-layer information transmission. Finally, the RexSE-Head (ResNeXt-Squeeze-Excitation) design is adopted to enhance the positioning accuracy of small defect targets.Results and discussion: The experimental results show that the mAP@0.5 of the improved model reaches 93.5%, and the number of parameters is only 3.29M, which realizes the high precision and high response performance of the detection of small defects in chemical special steels, and highlights the practical application value of the model. The code is available at https://github.com/improvment/prs-yolo.