Scale Adaptive Attention Network for Accurate Defect Detection From Metal Parts

Abstract

Metal component defect detection plays an important role in industrial manufacturing. However, it is a challenging task to detect defects from the metal component surface due to these problems: 1) Some defects are small and appear randomly on the metal component; 2) There is low-intensity contrast between defect areas and surrounding ones. To solve these issues, a Scale Adaptive Attention Network (SAA-Net) is proposed for defect detection from metal parts, where the Interactive Spatial Position Attention (ISPA) module is devised to detect small defects from the metal part surface by modeling the interdependence between pixels; then, the Dual Local-Global Transformer (DLGT) module is designed to distinguish the defect regions from the surrounding normal ones by fusing the overall attributes and key features. Experiments on the MPDD dataset demonstrate the effectiveness of the proposed SAA-Net, achieving the performance of 97.5%, 90.7%, and 96.1% on the pixel AUC, AP, and sPRO, respectively, further assisting in metal part detection in manufacturing.

Keywords

• Metal part monitoring
• defect detection
• attention mechanism
• feature extraction