Bidirectional 2D reservoir computing for image anomaly detection without any training

Abstract

Image anomaly detection is a crucial task in computer vision, where convolutional neural networks (CNN) often deliver exceptional performances. Hardware implementation of machine learning models is also important for achieving inference speed-up and power savings. However, the massive number of CNN parameters poses challenges for hardware implementation. This study introduces reservoir computing (RC) to create a compact image processor without training, thereby enabling scalable deployment. Our proposed bidirectional 2-dimensional reservoir computing (BiRC2D) is a feature extractor based on RC. Experiments conducted on the MVTec AD dataset, a benchmark dataset for real-world anomaly detection task, validated the efficacy of BiRC2D when integrated into the patch distribution modeling (PaDiM) framework. The mean intersection over union (mIoU) score from PaDiM with BiRC2D outperformed or was comparable to the mIoU score from PaDiM with ResNet-50 in several categories while reducing the parameter count by up to 98%.