TinySimpleNet: A SimpleNet for Unsupervised Industrial Anomaly Detection on Edge Devices
Abstract
Industrial anomaly detection is increasingly required in real-world inspection settings. Although recent methods achieve strong performance on standard benchmarks, many of them still rely on architectures that are difficult to efficiently deploy in real-world edge scenarios. Thus, we present TinySimpleNet, a compact reformulation of the SimpleNet architecture for industrial anomaly detection. The proposed model preserves the core idea of feature-space anomaly discrimination, while revisiting its architectural realization for low-complexity execution. Quantitatively, we evaluated the proposed approach on the MVTec-AD dataset and reported both image-level and pixel-level performance. Deployment experiments were conducted using the new STM32N6, a modern MCU platform featuring an integrated Neural-ART Neural Processing Unit (NPU) capable of delivering up to 600 GOPS for INT8 inference. The results show that a substantially more compact SimpleNet-style architecture (2.2 MB of SRAM at 18 FPS) can still retain meaningful anomaly detection capability, making it a promising direction for deployment-oriented industrial inspection.