WA-YOLO: A Wavelet-Transform-Enhanced Attention YOLO for Car Paint Defect Detection

Surface quality has become a key factor in vehicle manufacturing, making automatic paint defect detection increasingly essential. This study proposed a wavelet-transform-enhanced attention framework (WA-YOLO) to improve defect detection on car paint surfaces. The effectiveness of the wavelet transform for absolute phase map feature extraction was verified through supervised CNN evaluation, and the optimal wavelet basis was determined using both unsupervised clustering and supervised methods. Two novel attention modules, WT-SENet and WT-CBAM, were developed by adding discrete wavelet transform to analyze feature maps in frequency domains. By combining both modules, the proposed WA-YOLO achieved the highest accuracy, increasing the mAP@0.5 from 0.870 to 0.883 and the average class recall from 0.842 to 0.877. Notably, the recall for defect categories of low contrast, such as run, bulge, condensate, increased by 5 %, 3 %, and 15 %, respectively. These findings demonstrate the effectiveness of embedding frequency-domain information into attention mechanisms.