Scholarlink

<p>Surface defect detection is crucial for ensuring the quality of industrial products. However, due to the variability in defect sizes, types, and complex background textures, accurate in situ defect detection remains challenging. This study introduces a novel industrial surface defect detection network that leverages block-wise feature fusion. We propose the Block-wise Feature Fusion Convolution (BFFConv) to construct lightweight residual blocks, which generate multi-level features by segmenting the upper feature map and combining them with multi-scale convolution. This approach fully exploits local feature information, improving the model's judgment capabilities. Furthermore, we design a Multidimensional Convolutional Depooling Attention (MCDA) module that independently extracts local and global features through an adaptive pooling strategy and integrates them via well-designed convolutional layers. This enhances the model's ability to comprehend complex visual scenes. To facilitate training, we incorporate two Auxiliary Training Heads (AuxHead) into the network's middle layer, guided by auxiliary losses, to capture richer gradient information. Experiments on three benchmark datasets show that our proposed method effectively improves defect detection accuracy, achieving a mean Average Precision (mAP@0.5) of 77.7% on the NEU-DET dataset, 99.4% on the DAGM2007 dataset, and 95.4% on the PCB-DET dataset. The model also demonstrates scalability and robust performance, making it suitable for various industrial applications. The code and dataset required to reproduce these findings are available in the following GitHub repository: <a href="https://github.com/jh2000-code/bma-yolo.git">https://github.com/jh2000-code/bma-yolo.git</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Block-wise feature fusion for high-precision industrial surface defect detection

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Zhong Xiang,
  • Jianhui Jia,
  • Kailong Zhou,
  • Miao Qian,
  • Weitao Wu

摘要

Surface defect detection is crucial for ensuring the quality of industrial products. However, due to the variability in defect sizes, types, and complex background textures, accurate in situ defect detection remains challenging. This study introduces a novel industrial surface defect detection network that leverages block-wise feature fusion. We propose the Block-wise Feature Fusion Convolution (BFFConv) to construct lightweight residual blocks, which generate multi-level features by segmenting the upper feature map and combining them with multi-scale convolution. This approach fully exploits local feature information, improving the model's judgment capabilities. Furthermore, we design a Multidimensional Convolutional Depooling Attention (MCDA) module that independently extracts local and global features through an adaptive pooling strategy and integrates them via well-designed convolutional layers. This enhances the model's ability to comprehend complex visual scenes. To facilitate training, we incorporate two Auxiliary Training Heads (AuxHead) into the network's middle layer, guided by auxiliary losses, to capture richer gradient information. Experiments on three benchmark datasets show that our proposed method effectively improves defect detection accuracy, achieving a mean Average Precision (mAP@0.5) of 77.7% on the NEU-DET dataset, 99.4% on the DAGM2007 dataset, and 95.4% on the PCB-DET dataset. The model also demonstrates scalability and robust performance, making it suitable for various industrial applications. The code and dataset required to reproduce these findings are available in the following GitHub repository: https://github.com/jh2000-code/bma-yolo.git.