Scholarlink

<p>Industrial product surface defect detection plays a crucial role in guaranteeing the quality of industrial products. Several primary challenges remain in achieving efficient and accurate automatic detection: (1) Industrial product surface defects demonstrate irregular morphological features. (2) Significant size disparities between different defects result in the risk of information loss during feature fusion in Feature Pyramid Networks (FPN) typically employed in the neck component. (3) Visual similarity among different defect categories means that slight deviations in predicted box locations can result in misclassification errors. This research initially employs a DCN-C3k2 module design, incorporating DCNv3 to improve the model’s sensitivity to varied morphological information of objects. Furthermore, we design an Adaptive Focusing Diffusion Network (AFDN) that aggregates multi-scale features and implements adaptive channel selection and fusion, then propagates critical features through upsampling and downsampling processes to improve defect detection precision. Lastly, we introduce a Task Dynamic Interactive Detection Head (TDIDH). The TDIDH constructs detection head architecture according to the specific properties and variations of different detection tasks, with the objective of optimizing detection performance through enhanced inter-task dynamic interaction. Experiments are performed on public datasets GC10-DET and NEU-DET. The proposed method achieves AP, AP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> </InlineEquation>, and AP<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(_{75}\)</EquationSource> </InlineEquation> scores of 41.7%, 77.1%, and 40.8% respectively on NEU-DET, demonstrating improvements of 3.7%, 4.5%, and 4.1% compared to YOLOv11. Results on GC10-DET show scores of 44.3%, 81.5%, and 39.8%, with improvements of 3.3%, 0.9%, and 3.3% respectively compared to YOLOv11. Additionally, we achieve a 50% reduction in both parameter count and computational load through model pruning while preserving detection accuracy equivalent to the original AFFNet model, facilitating deployment on edge devices. The experimental results validate the effectiveness of our proposed method. The code is released at: <a href="https://github.com/yifansdut/affnet">https://github.com/yifansdut/affnet</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

AFFNet: Adaptive feature fusion network for defect detection of industrial product surface

  • 期刊论文

  • 发表日期：

    2026年1月24日

查看全文
  • Zhicheng Jia,
  • Shaoqing Wang,
  • Jinghua Zheng,
  • Xiaobo Han,
  • Yongwei Tang,
  • Fuzhen Sun

摘要

Industrial product surface defect detection plays a crucial role in guaranteeing the quality of industrial products. Several primary challenges remain in achieving efficient and accurate automatic detection: (1) Industrial product surface defects demonstrate irregular morphological features. (2) Significant size disparities between different defects result in the risk of information loss during feature fusion in Feature Pyramid Networks (FPN) typically employed in the neck component. (3) Visual similarity among different defect categories means that slight deviations in predicted box locations can result in misclassification errors. This research initially employs a DCN-C3k2 module design, incorporating DCNv3 to improve the model’s sensitivity to varied morphological information of objects. Furthermore, we design an Adaptive Focusing Diffusion Network (AFDN) that aggregates multi-scale features and implements adaptive channel selection and fusion, then propagates critical features through upsampling and downsampling processes to improve defect detection precision. Lastly, we introduce a Task Dynamic Interactive Detection Head (TDIDH). The TDIDH constructs detection head architecture according to the specific properties and variations of different detection tasks, with the objective of optimizing detection performance through enhanced inter-task dynamic interaction. Experiments are performed on public datasets GC10-DET and NEU-DET. The proposed method achieves AP, AP \(_{50}\) , and AP \(_{75}\) scores of 41.7%, 77.1%, and 40.8% respectively on NEU-DET, demonstrating improvements of 3.7%, 4.5%, and 4.1% compared to YOLOv11. Results on GC10-DET show scores of 44.3%, 81.5%, and 39.8%, with improvements of 3.3%, 0.9%, and 3.3% respectively compared to YOLOv11. Additionally, we achieve a 50% reduction in both parameter count and computational load through model pruning while preserving detection accuracy equivalent to the original AFFNet model, facilitating deployment on edge devices. The experimental results validate the effectiveness of our proposed method. The code is released at: https://github.com/yifansdut/affnet.