Novel variant transformer-based method for aluminum profile surface defect detection

Abstract The detection of surface defects on metals plays a pivotal role in the evaluation of aluminum profile product quality. To date, the most prevalent approach for detecting metal surface defects is via the use of conventional neural networks (CNNs). However, in recent years, transformer-based methods have achieved notable improvements in computer vision and exhibited superiority over traditional CNNs in most tasks. Regarding aluminum surface defect image characteristics, such as intricate textures, few defect samples, and large-scale differences in various defects, it is difficult to further improve the performance of traditional CNNs. Hence, this study proposes a novel variant transformer-based method for aluminum profile surface defect detection that combines a traditional CNN and transformer architecture with a deformable attention module (DAM). The DAM focuses on a subset of the critical sampling points around the reference point to alleviate the intractable complexity associated with high-resolution feature maps and enhances the recognition effect of small target defects. The proposed image-stitching method uses defect-free samples to generate new samples, which partially mitigates the issue of class imbalance. Moreover, we performed supplementary experiments to confirm the effectiveness of transfer learning in training small-scale datasets. Compared with a baseline, the mean average precision (mAP) improved by 3.32%. Extensive experimental results demonstrate the efficacy of our method in accurately detecting surface defects in aluminum profiles and significantly improving the detection of small target defects.