2026 年 17 巻 1 号 p. 93-107
We enhance bidirectional 2-dimensional (2D) reservoir computing (BiRC2D) by incorporating hierarchical pooling operations for computer vision tasks on edge devices. BiRC2D effectively captures local spatial dependencies in image data while being based on reservoir computing. However, BiRC2D lacks downsampling capabilities, limiting the ability to capture multi-scale image structures. To address this limitation, we introduce a hierarchical extension by alternately stacking BiRC2D blocks and 2D pooling layers. This enhancement enables progressive spatial feature abstraction while preserving the low-parameter, training-free advantages of reservoir computing. Anomaly detection experiments on the MVTec AD dataset demonstrate that feature embedding-based methods using our proposed architecture achieve competitive performance while reducing the parameter count by 97-99% compared to those using ResNet-50. Our proposed architecture operates solely through random spatial dynamics, offering efficient and scalable anomaly detection. These properties make it particularly well-suited for energy-constrained, real-time industrial inspection systems.
