Abstract
Image anomaly detection (IAD) is an important computer vision task for industrial man- ufacturing applications like battery surface anomaly detection, ceramic defect detection, food inspection, and so on. In the computer vision community, there is a primary focus on unsupervised methods, but few of them are used in real manufacturing. Because there is little analysis of the industry’s needs, i.e, the number of training images are limited. Because of this, it is important and urgent to design a few-shot IAD algorithm to adapt the demands of industrial manufacturing. Accordingly, this thesis aims to fill the above gaps.
To begin with, we propose a uniform benchmark to assess the capability of the existing IAD algorithms, which includes several aspects, i.e., few-shot learning, continual learning, noisy labels, various levels of supervision, memory usage, and inference speed. Specifically, we skillfully build a comprehensive image anomaly detection benchmark (IM-IAD) with 16 algorithms on 7 mainstream datasets with few-shot anomaly detection scenarios. Our ex- tensive experiments (17,017 in total) provide in-depth insights for few-shot IAD algorithm redesign or selection. Next, the proposed benchmark IM-IAD gives challenges as well as directions for the future, especially for few-shot IAD.
After that, a new feature representation method has been proposed to address the demands of manufacturing-based few-shot IAD. We reveal that the rotation-invariant feature has a significant impact on few-shot IAD. Hence, we construct a graph-based model (GraphCore) to provide a novel visual isometric invariant feature (VIIF) as an anomaly measurement feature. As a result, VIIF can robustly improve the anomaly-discriminating ability and can further remove redundant features by a significant amount. Extensive experiments on MVTec AD and MPDD verify the effectiveness of GraphCore.
Furthermore, we tackle the problem of few-shot IAD in the changeover procedure. We introduce the COAD dataset, the first comprehensive dataset for this task. Specifically, the CO-AD dataset contains similar objects spanning 21 domains, multiple anomalies (up to 4) with different locations in a single view, and multiple defect types (up to 5) for each object. Moreover, we reveal that existing few-shot IAD methods do not perform well in COAD. To this end, we propose a simple but effective anomaly transfer method(Transfer- AD), which fully utilizes the anomalies of the source domain dataset by copy-pasting them into the normal dataset of the target domain. Comprehensive experiments on the CO-AD dataset clearly demonstrate the effectiveness of Transfer-AD.