Deep Neural Networks for Image Anomaly Detection: Application in Real World Industrial Scenarios

Deep neural network is the new norm in the present era, where it’s being used in almost all the evolving fields, so is the field of anomaly detection. With the modern IT infrastructure industries are in a constant quest to search for new algorithms to analyze the data in order to be more autonomous, agile, efficient, and cost-effective. Anomaly detection is one such task, which industries want to automate as it finds its application in various fields like banking, traffic management, manufacturing, online fraud detection, anomalous behavior, etc. In this dissertation, we explored various novel approaches to solve the image anomaly detection problem. We proposed ways for anomaly detection tasks, keeping real industrial problems in mind like data scarcity, imbalanced data, limited resources, etc. The dissertation proposes novel models like adapted capsnet, stacked capsule autoencoders, Pyramidal Image Anomaly Detector (PIADE), and Vision Transformer for Image Anomaly Detection and Localization (VT-ADL) for Deep Anomaly Detection (DAD) task. The methods can be trained in both supervised and semi-supervised ways. And VT-ADL is capable of global image classification and localization, without any need for pixel-precise ground truth data. We tested our methods on various academically used datasets like MNIST, CIFAR, COIL, FMNIST, and real industrial datasets like MVTec and BTAD. The proposed methods performed at par or better than the state-of-the-art methods.