Factor Graph-Optimized Cross-Scale Detector for Wafer Defect Detection

Abstract

Accurate wafer defect detection is critical for yield improvement in semiconductor manufacturing and package testing. Conventional CNN-based and Transformer-based detectors struggle to localize small defects and capture global context in low-resolution images with extreme scale variation. In this paper, we propose FGC-DETR, a factor graph-optimized cross-scale detector tailored for wafer defects. FGC-DETR harnesses factor graph optimization to align cross-scale features and integrates a Laplacian edge-enhancement module to amplify micro-crack signals, enabling efficient and robust detection across extreme scale ranges. Evaluated on a real-world dataset of custom wafer images, FGC-DETR achieves 95.9% AP@0.5 and 12.0% FNR at 10.49 M parameters, outperforming mainstream methods and demonstrating its practical value for real-time production deployment in AOI systems. Code is available at: https://github.com/mindednessr/FGC-DETR.