Enhancing Anomaly Detection Performance and Acceleration

Abstract

Automation of visual inspection is a critical aspect in industrial fields. Recently, research on anomaly detection using neural networks has been gaining increasing attention. In particular, approaches that use a pre-trained convolutional neural network have exhibited high performance. In this study, we focused on PatchCore, which is a high-performance model, and further improved it using two high-resolution images to accurately detect small anomalies. However, the extracted features (memory bank) consume a large amount of memory and storage; the memory bank is compressed by k-means clustering. Moreover, the inference time was reduced by an approximate nearest-neighbor search using an inverted index. Our method achieved an image-level AUROC of 0.994 on the MVTec anomaly detection dataset. In addition, a pixel-level AUROC of 0.984 was achieved, which is better than that of PatchCore. Furthermore, the compression time was reduced by more than 97% by clustering the memory bank using k-means while maintaining the performance.