BIIR: Blind Inpainting based Image Reconstructon for Texture Defect Detection

Image reconstruction is an important method for texture defect detection, and the existing image reconstruction algorithms based on Autoencoder and GAN cannot suppress the reconstruction of defect information, which affects the detection accuracy. To solve this problem, this paper proposes a novel image reconstruction algorithm based on image inpainting, which includes two modules of defect estimation network and defect inpainting network. Firstly, the defect estimation network used the pre-training model to extract the deep features of the defect image and applied the Gaussian distance to estimate the background area. and then the image inpainting network applied the contextual attention mechanism to repair the non-background area of the defect image. Through the experimental analysis which compared with other state-of-the-art image reconstruction algorithms on the public Mvtec texture dataset, the superiority of the proposed algorithm is effectively verified.