About the Application of Autoencoders For Visual Defect Detection

Visual defect detection is a key technology in modern industrial manufacturing systems. There are many possible appearances of product defects, including distortions color, shape, contamination, missing or superfluous parts. For the those, besides traditional image processing techniques, convolutional neural networks based methods have also appeared to avoid usage hand-crafted features and build more efficient mechanisms. In our article we deal with autoencoder (AEs) which do not require examples defects for training. Unfortunately, manual and/or trial-and-error design AEs still required achieve good performance, since there unknown parameters can greatly influence abilities. study chosen well performing AE known as structural similarity (SSIM-AE), where loss function comparison output input implemented via SSIM instead often used L1 L2 norms. Investigating performance SSIM-AE on different data-sets, found that its be improved modified structures without modifying size latent space. We show finding model low reconstruction error during training does mean abilities denoising increase efficiency.