RSDD: A Rail Surface Defect Generation Model Based on Adaptive Semantic Positional Decoupled Diffusion

Rail surface defect detection is of great significance in the field of rail traffic safety, and rail damage detection algorithms based on deep learning are increasingly widely used. However, it is difficult to collect rail damage samples, and the number of available damage samples is small, which cannot provide sufficient feature information for the detection algorithm. The existing defect generation algorithms produce samples with poor authenticity and lack of diversity, and the generated defect images are often difficult to accurately match with the corresponding defect masks. To address this problem, this paper proposes a diffusion model-based generation method RSDD (Rail Surface Detcet Difussion). It decouples the defect information into a defect embedding characterising the appearance information and a spatial embedding characterising the location information, and is able to generate different types of defects at a specified location. In addition, RSDD designs an adaptive attention weight redistribution mechanism, which is able to adjust the cross-attention based on the difference between normal samples and generated images in the process of diffusion inference, so as to improve the matching between defective images and defective masks. Experiments demonstrate that RSDD is able to generate high-quality and rich defect images, significantly facilitating downstream rail damage detection.