Evaluation of Generative Neural Networks for Automatic Defect Detection

Quality assurance of mass produced items is prone to errors when performedmanually by a human. This has created a need for an automated solution. Theemergence of deep neural networks has created systems that can be trained toclassify defect from non-defect items. However, to alleviate the need for largeamounts of manual labeling required for most classification networks, severalunsupervised methods have been used. This report evaluates the use of a deepautoencoder for unsupervised defect detection. Furthermore is the use of anautoencoder compared to applying inpainting and a generate adversarialnetwork(GAN) for the same task.The report finds that the autoencoder used could find the largest of defects testedbut not the smaller ones. It is also shown that neither use of inpainting nor a GANimproved on the autoencoder result. It is of note however that it was a naiveimplementation of inpainting and the GAN and they were lacking some state of theart aspects