Latent normal images-based zero-negative sample rail surface defect segmentation method

Rail surface defects pose a significant risk to the safe operation of railways, making rapid and accurate detection essential. However, existing deep learning-based detection methods struggle to identify all potential defects that may occur during operation due to imbalanced sample, which limits practical application in railway maintenance. To address this, a latent normal images-based zero-negative sample segmentation method for rail surface defects is proposed. This method utilizes an improved Pix2Pix network, which learns the characteristics of normal rails to generate latent normal images. Defect regions are then inferred based on the differences between input detection image and latent normal image. Experimental results demonstrate that the proposed method achieves a mPA of 0.9984 and a mIoU of 0.8305. Under the zero-negative sample condition, it performs comparably to other classical segmentation models, such as DeepLabv3+ and U-Net3+, which rely on a large number of labeled negative samples. Additionally, the proposed method shows better adaptability to new defects.