Scholarlink

<p>To enhance the performance of industrial defect detection and address challenges related to high inter-class similarity and significant intra-class variation, this paper proposes a novel deep learning model, DMSKan-attenNet. The model integrates multi-scale KAN convolution, depthwise separable convolution, and the GSHilo Attention mechanism to improve feature extraction, computational efficiency, and multi-level feature fusion. First, a multi-scale KAN convolution structure is designed to simultaneously capture local details and global features by employing KAN convolution kernels of varying sizes in parallel. Additionally, an interaction matrix is introduced to model correlations between different scales, enhancing feature fusion accuracy. Second, depthwise separable convolution is incorporated to reduce computational complexity while preserving feature extraction capabilities, thereby improving the model's adaptability. Finally, we propose the GSHilo Attention mechanism, which integrates spectral attention and a gated fusion module to refine high- and low-frequency feature fusion, thereby enhancing the model’s generalization ability for complex industrial defects. Extensive validation experiments conducted on the NEU-DET and GC10-DET datasets demonstrate that DMSKan-attenNet achieves superior feature extraction and generalization performance, significantly improving industrial defect detection accuracy.</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

Enhanced Industrial Defect Detection with Multi-scale KAN Convolution and GSHilo Attention Mechanism

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Xin Zhang,
  • Ruocheng Wang

摘要

To enhance the performance of industrial defect detection and address challenges related to high inter-class similarity and significant intra-class variation, this paper proposes a novel deep learning model, DMSKan-attenNet. The model integrates multi-scale KAN convolution, depthwise separable convolution, and the GSHilo Attention mechanism to improve feature extraction, computational efficiency, and multi-level feature fusion. First, a multi-scale KAN convolution structure is designed to simultaneously capture local details and global features by employing KAN convolution kernels of varying sizes in parallel. Additionally, an interaction matrix is introduced to model correlations between different scales, enhancing feature fusion accuracy. Second, depthwise separable convolution is incorporated to reduce computational complexity while preserving feature extraction capabilities, thereby improving the model's adaptability. Finally, we propose the GSHilo Attention mechanism, which integrates spectral attention and a gated fusion module to refine high- and low-frequency feature fusion, thereby enhancing the model’s generalization ability for complex industrial defects. Extensive validation experiments conducted on the NEU-DET and GC10-DET datasets demonstrate that DMSKan-attenNet achieves superior feature extraction and generalization performance, significantly improving industrial defect detection accuracy.