中国科学院自动化研究所机构知识库

In this paper, we propose an unsupervised background reconstruction method to detect defects on surfaces with unevenly distributed textures. An improved deep convolutional autoencoder is utilized to reconstruct the textured background of the original image as a defect-free reference. Specifically, a weighted loss function based on structural similarity (SSIM) is utilized to adapt to the unevenly distributed texture background and improve the reconstruction accuracy. Furthermore, combined with the reconstructed defect-free reference, a novel difference analysis method based on the discrete cosine transform (DCT) is given to accurately segment the defect regions from the original image. A series of experiments for the defect detection on mobile phone cover glass (MPCG) are conducted. The processing time for an image of 512 x 512 pixels is only 20 ms, which satisfies the requirement of online detection. The experimental results verify the effectiveness of the proposed method.