AE5-SSIM: A Novel Unsupervised Tinfoils Defect Detection Model with Deep Autoencoder

Tinfoils defect detection plays an important role in the field of industrial manufacturing. Human visual evaluation is not only time-consuming and labor-consuming but also subjective and error-prone. There were some defect detection algorithms with deep learning proposed to deal with this problem. However, most of the existing algorithms need to be trained by large numbers of normal and abnormal training samples. It is difficult to collect and mark a lot of abnormal samples, thus limiting the usage of these methods. In this paper, we propose a novel unsupervised tinfoils defect detection model with deep autoencoder, called AE
5
-SSIM (structural similarity-based reconstruction autoencoder, whose encoder and decoder networks have five sub-encoder blocks and five sub-decoder blocks, respectively), to break the constraints of limited abnormal training samples on algorithms. We only use the normal areas of tinfoils samples to train the AE
5
-SSIM model. After trained, this model can reconstruct the input tinfoils images to normal images. Finally, combining the input image and the reconstructed image, a classifier is designed to classify the defects. Experimental results on our public dataset proved the efficiency and effectiveness of the proposed algorithm, which achieved 95.33% AUC for the tinfoils defect detection.