A Lightweight Conditional Diffusion Segmentation Network Based on Deformable Convolution for Surface Defect Detection

Surface defect detection is crucial to industrial manufacturing and research for surface defects has drawn much attention. However, defects in industrial environment are very diverse. Because defects scale and poses are constantly changing and current methods lack the ability to model the deformation. To solve this problem, a lightweight conditional diffusion segmentation network based on deformable convolution is proposed. First, the conditional diffusion process is introduced for effective feature extraction; by gradually corrupting the defect images and recovering them from latent space, the model can obtain pixel‐level segmentation results in an iterative process. Second, the efficient feature extraction block is proposed to address the problem of modeling varying defects, which is designed with a partial deformable convolutional layer that can fully extract geometric features of the diverse defects to further enhance the modeling power of the proposed network. Furthermore, the hyperparameters of the diffusion process are discussed to further improve the performance of the proposed method. The experimental results on DAGM2007, MT, AeBAD, and MVTec‐AD indicate that the proposed model performs better than other baseline models.