Unsupervised Defect Detection based on Boundary Equilibrium Generative Adversarial Network

In addition to the brand and sales channels that can be found everywhere in the shoe industry, the foundry manufacturing industry is also an important part of the industrial chain. The traditional shoe industry has high manpower requirements but a low ratio of information personnel and high-tech equipment. Many testing procedures can only be performed manually. The complexity of footwear products is high, and the entire inspection process depends on a large amount of manpower. Therefore, problems such as consuming too much manpower, lack of efficiency, difficulty in precise inspection, quality requirements that vary from person to person, and incomplete quality management are extended. Through traditional machine learning methods, most of them are supervised learning methods. Under this method, a large number of negative samples should be always collected. It is very difficult to collect these negative samples in actual industrial production. Therefore, this article proposes a defect detection model based on unsupervised learning. As long as there are enough positive samples to be trained, we use the BEGAN model to modify it and combine another Autoencoder. This model is much easier and faster to be trained than the traditional GAN model, better responding to the footwear industry with more product types.