Scratch Defects Detection of Curved Metal Surface Based on Multiple High Frequency Projection and Inverse Transfer Learning

The accurate detection of scratch defects on curved metal surfaces remains challenging in machine vision-based defect detection. Compared with normal metal flat surfaces, curved metal surfaces often exhibit unavoidable specular halo, which can obscure defect features. Additionally, in real-world scenarios, collecting and labeling scratch defect data is challenging and exhibits some stylistic differences from publicly available scratch datasets. Consequently, neural network models trained on public datasets demonstrate low generalization performance. The article proposes to tackle this issue from two aspects. First, we use a multiple high-frequency projection illumination imaging method to generate high-contrast images of metal part surfaces, which can eliminate the interference of highlights. Meanwhile, we used the above method to build a dataset reflective metal surface scratch (RMSSC) containing 50 images of RMSSC. Second, we build a defect detection neural network and improve its generalization performance using inverse transfer learning. Unlike traditional transfer learning, inverse transfer learning does not update the detection network parameters but instead modifies the high-level semantic representation of the defects being detected. The proposed method is evaluated on the public datasets NEUSD-900 and RMSSC both qualitatively and quantitatively. The results demonstrate that our method outperforms state-of-the-art approaches by more than 19.65% in mPA and 18.35% in mIoU.