LASDNet: A Lightweight Adaptive Surface Defect Detection Network

Qiancheng Wei, Yi Yuan, Ying Liu, Mohammed Ali Mohammed Al-Hababi, Qiya Su, Muyao Yu

Abstract: Surface defect detection is an important task in industry. However, surface defect detection still faces many challenges, including variations in aspect ratios, similarity with the background, and difficulty in detecting small defects. In this paper, we propose LASDNet, a novel model for defect detection. LASDNet generates predictions for three different categories: the center point of the defect, the offset of the center point, and the size of the defect. The location of the defect is determined by its center point and size, with the center point being adjusted by the offset. Firstly, an adaptive module is designed to generate the ground truth heatmap based on the shape of the defect, thereby enhancing the precision of the ground truth. Secondly, the hourglass backbone is optimized by redesigning the structure to enhance its capacity for detecting small defects. Finally, an intermediate supervision module is proposed to utilize multi-scale features to locate the defect from rough to precise. Our proposed LASDNet model outperformed all its peers on the public rail defect dataset (RRTD) and shot-circuit defects of printed circuit boards dataset (PCB).