Large Visual Perception Network with Auxiliary Supervision for Surface Defect Detection

Surface defect detection is a pivotal process in industrial manufacturing. However, the multi-scale and the low-contrast problems are still difficult to overcome in current defect detection methods. This article proposes a large visual perception network with auxiliary supervision to address these issues. Specifically, the large visual perception network is designed to enhance the robustness of the model treating the defects with multi-scale changes. In addition, a deformable convolutional feature alignment module is proposed to redesign the auxiliary supervision architecture, which provides more accurate gradient information for the backbone network. The effectiveness of this model is evaluated through experimentation on the four datasets. The results illustrate that the proposed methodology outperforms current state-of-the-art systems on all datasets, thus affirming its potential to advance the field of surface defect detection.