A Multiscale Convolutional Registration Network for Defect Inspection on Periodic Lace Surfaces

Defect inspection on lace products has always been a challenging task. The complexity and fragility of lace cloths increase the difficulty in distinguishing defects from normal textures. It is also extremely difficult to collect enough defective samples to support the training of supervised learning models. In this article, we propose a detection approach to inspect and locate defects on periodic lace surfaces, which only requires defect-free image samples. The approach consists of three steps: extract image patches and their corresponding defect-free patches, reconstruct contrast patches to increase the morphological similarity of input image pairs, inspect defects on the residual map between the output patch and the image patch to be tested. This method has two prominent advantages: completely unsupervised and lightweight. Experiments are conducted to evaluate the performance of the framework, of which the results confirm the effectiveness and superiority of the proposed model compared to the baseline model.