Developing and Experimenting Deep Learning Methods for Unsupervised Anomaly Detection in Images

Anomaly detection in the industrial sector is an important problem as it is a key component of quality control systems that minimize the chance to miss a defective product. Most often, the anomaly detection is done through analysis of the images of the products. Because the products, or their designs change and quality data is hard to obtain, this problem is approached in an unsupervised manner. There are many different anomaly detection approaches, but most of them deal with low dimensional data and do not work well with the images. We examine deep learning techniques that utilize convolutional neural networks which can extract meaningful image representations to a lower-dimensional space. It allows the models to learn the important features of an image, regardless of some small changes in the input. The feature extracted with the CNN are used to train standard one class classifier (such as one class support vector machine) that is able to classify an object in an unsupervised way. Then next approach is an anomaly detection based on generative adversarial networks (GAN). The network learns a mapping from the latent space to a representation of a “normal” data and is able to produce new and unseen data samples from random latent vectors. In particular, the main state of the art models for anomaly detection based on GAN were examined. Afterwards has been developed a new model (based on BiGAN) to detect anomalies to remove the problems of the standard methods and increase the performance called CBiGAN