Progressive Feature Enhancement Network for Surface Defect Segmentation

Surface defect detection is critical for maintaining the high quality of industrial products. However, defect detection is confronted with challenges, such as diverse defect types and scales, low contrast, and complex backgrounds. To tackle the problems, we propose a Progressive Feature Enhancement Network (PFENet), which aims to gradually strengthen the representation of features through semantic-guided Single-scale Feature Enhancement (SFE) module and Cross-scale Feature Enhancement (CFE) module. Specifically, SFE highlights defect semantic information of multi-level features by exploiting spatial similarities between the features and high-level features. CFE adaptively selects important defect information and suppresses redundant information through the mutual interaction of cross-level features. The mutual interaction enlarges the difference between foreground and background and facilitates learning more discriminative defect features for complex defects. Extensive experiments on three publicly available defect datasets, magnetic tile (MT), NEU-Seg, and Road defect dataset demonstrate that the proposed method achieves state-of-the-art performance.