Small-object feature restoration-based defect detection of flexible printed circuits

Abstract Flexible printed circuits (FPCs) are extensively employed in consumer electronics, automotive systems, and a variety of other domains. During the manufacturing process, surface defects—including scratches, indentations, and ink contamination—commonly arise and are typically characterized by small size, indistinct boundaries, and low contrast. These defects are frequently embedded within complex textured backgrounds, which render them easily overlooked during feature extraction and presents substantial challenges for automated inspection. To address these challenges, we propose YOLOX-GSM, a GAN-based framework specifically designed for small-object defect detection in FPCs. The framework integrates a Small-object feature information restoration module, which recovers critical details lost during downsampling, and a progressive spatial-aware attention module that suppresses background interference. Furthermore, similarity regularization is incorporated into the discriminator to enforce consistency between generated and original features, while a small-object spatial-aware module is embedded into the backbone to strengthen the model’s focus on critical defect regions. Comprehensive evaluations were performed on a self-constructed FPC surface defect dataset (FPCSD2023), alongside public datasets TT-100K and GC10-DET. Experimental results demonstrate that on FPCSD2023, average precision for small objects ( mA P s m a l l ) increased by 5.4%, with mA P 50 improving by 8.4%. On TT-100K, these metrics improved by 5.0% and 10.5%, respectively. Moreover, on GC10-DET, the proposed method outperformed current state-of-the-art defect detection approaches, substantiating its efficacy and promising applicability in FPC small-object defect detection. Code available at https://github.com/FutureIAI/YOLOX-GSM.git .