EDSV-Net: An efficient defect segmentation network based on visual attention and visual perception

In industrial production, surface defect detection algorithms based on convolutional neural networks have been widely studied to improve production quality. However, for practical applications, there are still many issues to be solved, such as the complexity and diversity of defect categories, the difficulty of obtaining defect samples, and the difficulty of existing algorithms in accurately segmenting defects. To solve these issues, we present an effective defect segmentation network based on visual attention and visual perception termed EDSV-Net. Specifically, we use ResNet18 as the backbone network in EDSV-Net. Then a multi-scale feature extraction (MSFE) module is introduced to enhance the scale invariance of high-level features and the diversity of contextual features. In addition, a spatial attention (SA) model combined with a channel attention (CA) model is applied to low level features and MSFE features, respectively, to extract more effective spatial and semantic information. Moreover, a depthwise separable convolution is introduced to reduce the network complexity. Finally, due to the issues of existing defect detection algorithms ignoring structural similarity and defects being difficult to obtain, we design a balanced defect and structural measure loss function. Meanwhile, we propose a structural similarity measure, which combines the pixel similarity for evaluation. EDSV-Net only requires no more than 60 random abnormal samples to obtain accurate segmentation results and the real-time performance meets the requirements of actual industrial production. Based on three challenging real-world defect datasets, the results of the evaluation demonstrate that EDSV-Net outperforms seven state-of-the-art methods on accuracy and real-time performance.