Scholarlink

<p>At present, the detection of industrial surface defects often faces challenges such as difficult identification of complex background defects, weak semantic information contained in tiny defects and significant changes of defect objects, which are related to product quality and people’s life safety. To solve the above problems, this paper proposes a novel industrial defect detection method (WTCF-Net). Firstly, to prevent the loss of semantic information in the process of layer-by-layer transmission of feature information, we propose a Wavelet Feature Convolution block (WFC), which decomposes the image into low-frequency and high-frequency components of different scales through wavelet transform to enhance the representation of defect features and suppress redundant information. Secondly, to improve defect feature recognition under complex background, we design an Interactive Residual Module (IRM) to aggregate feature information of different dimensions. At the same time, the Interactive Residual Feature extractor (IRF) is used to capture cross-dimensional information interaction and establish the dependency relationship between different dimensions, so that the network can capture more subtle complex features. Then, to improve the detection of multi-scale defects in industrial products, this paper proposes a Cross-level Feature Aggregation Network (CFA-Net), and designs a three-layer gradually decreasing feature fusion path to enhance the interaction between different scales through the fusion of adjacent layers and cross-layers. In addition, the Feature Enhancement Module (FEM) is designed. By aggregating the features of adjacent levels, cross-level semantic association is established, thus improving the feature extraction ability of the model for multi-scale defects. At the same time, Collaborative Filtering Module (CFM) is used to filter redundant information after fusion, which enriches feature representation. Finally, we design the loss function of defect regression intersection ratio (CDIou), and improves the accuracy of defect detection by calculating the loss of width and height product. The validity and generalization of the model is tested on the data sets of NEU-DET, PCB and DeepPCB. Compared with the baseline model, the mAP value of the model is increased by 5.9%, 1.3% and 1.4%. The detection speed is 53 FPS, which has a wide application potential.</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

WTCFNet for industrial defect detection using wavelet transform and cross layer feature fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月23日

查看全文
  • Hao Chen,
  • Yu-Bo Ren

摘要

At present, the detection of industrial surface defects often faces challenges such as difficult identification of complex background defects, weak semantic information contained in tiny defects and significant changes of defect objects, which are related to product quality and people’s life safety. To solve the above problems, this paper proposes a novel industrial defect detection method (WTCF-Net). Firstly, to prevent the loss of semantic information in the process of layer-by-layer transmission of feature information, we propose a Wavelet Feature Convolution block (WFC), which decomposes the image into low-frequency and high-frequency components of different scales through wavelet transform to enhance the representation of defect features and suppress redundant information. Secondly, to improve defect feature recognition under complex background, we design an Interactive Residual Module (IRM) to aggregate feature information of different dimensions. At the same time, the Interactive Residual Feature extractor (IRF) is used to capture cross-dimensional information interaction and establish the dependency relationship between different dimensions, so that the network can capture more subtle complex features. Then, to improve the detection of multi-scale defects in industrial products, this paper proposes a Cross-level Feature Aggregation Network (CFA-Net), and designs a three-layer gradually decreasing feature fusion path to enhance the interaction between different scales through the fusion of adjacent layers and cross-layers. In addition, the Feature Enhancement Module (FEM) is designed. By aggregating the features of adjacent levels, cross-level semantic association is established, thus improving the feature extraction ability of the model for multi-scale defects. At the same time, Collaborative Filtering Module (CFM) is used to filter redundant information after fusion, which enriches feature representation. Finally, we design the loss function of defect regression intersection ratio (CDIou), and improves the accuracy of defect detection by calculating the loss of width and height product. The validity and generalization of the model is tested on the data sets of NEU-DET, PCB and DeepPCB. Compared with the baseline model, the mAP value of the model is increased by 5.9%, 1.3% and 1.4%. The detection speed is 53 FPS, which has a wide application potential.