Knowledge-Guided Neural Architecture Search for Anomaly Localization
Cheng Qian, Xiaoxian Lao, Chunguang Li · 2024
Anomaly localization is an important problem arising in various industrial scenarios. Nowadays, deep learning-based anomaly localization methods have made significant progress. Most of these methods employ manually designed network architectures. Neural architecture search (NAS) is a technique to automate the process of designing network architectures. If NAS is used to find an efficient architecture for anomaly localization, improvement in localization performance can be expected. To find a good network architecture, a validation set is needed to evaluate the performance of various architectures. However, usually only a small number of anomalous samples are available, making it hard to construct a good validation set, thus affecting the accuracy of performance evaluation and further the search efficiency. In many cases, we can obtain some knowledge of anomalies summarized by domain experts. Utilizing such knowledge can help us generate anomalous samples and build a good validation set. In this paper, we propose a knowledge-guided anomaly generator (KGAG) to generate anomalous samples. Besides, we weight the generated anomalies based on their anomaly-like degree and thus put forward a novel reward for reinforcement learning (RL)-based NAS. We conduct experiments on various datasets. The experimental results demonstrate the advantages of the proposed knowledge-guided RL-based NAS method.