Aspect-Ratio and Scale Aware Rotated Defect Detection for Fiber-reinforced Composites

Detecting surface defects in carbon fiber-reinforced composites (FRCs) during prepreg stacking is crucial for ensuring product quality. Traditional methods using horizontal bounding boxes (HBBs) for defect detection often lead to overlapping boxes and inaccurate results. This paper proposes an Aspect-Ratio and Scale Aware Detector network (ARSADet) specifically designed for rotated defect detection in FRCs. The network incorporates a Gabor convolutional network for feature extraction, an aspect-ratio and scale-aware label assignment method, and a newly established FRCs Oriented Defect Detection dataset (FODD). Evaluation on the FODD dataset demonstrates superior performance compared to existing methods.