Anomaly Detection in Textured Surfaces

Abstract

Detecting anomalies in textured surfaces is an important and interesting problem that has practical applications in industrial defect detection and infrastructure asset management with a lot of potential financial benefits. The main challenges in this task are that the definition of anomaly changes from domain to domain, even noise can differ from the normal data but should not be classified as an anomaly, lack of labelled datasets and a limited number of anomalous instances. In this research, we have explored weak supervision and network-based transfer learning for anomaly detection. We developed a technique called AnoNet, which is a novel and compact fully convolutional network architecture capable of learning to detect the actual shape of anomalies not only from weakly labelled data but also from a limited number of examples. It uses a unique filter bank initialization technique that allows faster training. For a HxWx1 input image, it outputs a HxWx1 segmentation mask and also generalises to similar anomaly detection tasks. AnoNet on an average across four challenging datasets achieved an impressive F1 Score and AUROC value of 0.98 and 0.94 respectively. The second approach involved the use of network-based transfer learning for anomaly detection using pre-trained CNN architectures. In this investigation, fixed feature extraction and full network fine tuning approaches were explored. Results on four challenging datasets showed that the full network fine tuning based approach gave promising results with an average F1 Score and AUROC values of 0.89 and 0.98 respectively. While we have successfully explored and developed a method each for anomaly detection with weak supervision and supervision from a limited number of samples, research potential exists in semi-supervised and unsupervised anomaly detection.