Surface Defect Anomaly Detection Based On Generative Adversarial Networks

As a classical topic in computer vision,anomaly detection has broad application prospects.Surface defect detection is a crucial procedure in industrial manufacture.A majority of traditional defect detection methods are based on machine vision or supervised learning paradigm.Massive supervised defect samples are needed when training those methods.However,it is difficult to get enough defect samples during manufacturing.Because of some unexpected factors,the model can't detect all kinds of defects in test stage.With the development of deep unsupervised learning,autoencoders or generative adversarial networks are widely used in surface defect detection problem,which are only based on modeling normal data to distinguish abnormal defects.Therefore,this paper proposes two defect detection methods based on generative adversarial networks.(1)Traditional autoencoders often fall into trouble that the reconstruction ability is too high,that's to say the models can lead to undesirable reconstruction of abnormal or defective parts.It is harmful to make right classification decisions.This work utilizes a mechanism to restrain latent code,which maintains consistency of the encoding of the input image and reconstruction encoding in feather space.Self-attention and hierarchical decoding in Big GAN are used to enhance the ability of feather learning,in addition,generative adversarial learning improves the reconstruction performance of the model.(2)The structures of reconstruction networks are always designed by experts,while neural architecture search contributes to find proper network architectures which are suitable for some tasks and have good performance.This work combines neural architecture search and generative adversarial networks to automatically search encoder-decoder architectures by using gradient-based search algorithms.Finally,this work evaluates the two mentioned methods on the MVTec which contains 15 different types of industrial inspection image datasets.Though evaluation metric AUROC,the experimental results using quantitative and qualitative analysis demonstrate the effectiveness of the proposed methods.