Zhejiang Daxue xuebao. Lixue ban (Jan 2025)
A human visual cognitive mechanism based network for surface defect detection(基于人类视觉认知机制的表面缺陷检测)
Abstract
Surface defect detection is crucial for ensuring product performance, quality, aesthetics, and production efficiency. Despite rapid advancements in artificial intelligence in the field of visual inspection, developing machine vision learning based on biological visual cognition remains a challenging issue requiring in-depth research. This paper proposes a surface defect detection network based on human visual cognitive mechanism (HVCM-Net). Specifically, at the macro level, by simulating the working principles of the foveal and parafoveal areas on the retina, a parallel backbone network consisting of a foveal vision branch and a peripheral vision branch is proposed, learning high spatial frequency local details and low spatial frequency global semantics from defective images. At the micro level, the dynamic weighting fusion module (DWFM) dynamically integrates the output feature maps of the two branches in an adaptive manner, enabling the model to learn and filter more comprehensive, accurate, and complementary defect features. Additionally, the fusion branch introduces the feature preserving downsampling (FPD) module that employs patch merging technique, effectively alleviating the issue of minor defect information loss caused by traditional downsampling. HVCM-Net achieves superior detection performance on defect datasets GB-DET, NEU-DET, and DAGM2007 compared to other algorithms, demonstrating the effectiveness of the proposed method.(进行表面缺陷检测是确保产品性能、质量、美观度以及生产效率的重要手段。尽管人工智能在视觉检测领域取得了飞速发展，但基于生物视觉认知指导机器视觉学习的方法，仍是研究难点。提出了一种基于人类视觉认知机制的表面缺陷检测网络（HVCM-Net）。在宏观层面，模拟视网膜上中央凹和中央凹外区域的工作原理，提出了中央视觉分支和外周视觉分支并行的骨干网络，分别负责学习缺陷图像的高空间频率局部细节信息和低空间频率全局语义信息。在微观层面，动态权重融合模块（DWFM）以自适应的方式融合两个分支的输出特征图，可学习和过滤更全面、准确和互补的缺陷特征。另外，融合分支引入特征保存下采样（FPD）模块，采用特征拼接技术，有效缓解了传统采样可能产生的微弱缺陷信息丢失问题。HVCM-Net在缺陷数据集GB-DET、NEU-DET和DAGM2007上取得了优于其他方法的检测性能，验证了其有效性。)
Keywords
