TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection

TL;DR

TinyGLASS is a lightweight, real-time, self-supervised anomaly detection system optimized for in-sensor processing on resource-limited edge devices, achieving high accuracy and efficiency.

Contribution

It introduces TinyGLASS, a compact adaptation of GLASS with deployment optimizations for in-sensor platforms, enabling real-time anomaly detection with significantly reduced model size.

Findings

Achieves 8.6x parameter compression with competitive accuracy.

Operates at 20 FPS within 8 MB memory constraints.

Demonstrates low power consumption and high energy efficiency.

Abstract

Anomaly detection plays a key role in industrial quality control, where defects must be identified despite the scarcity of labeled faulty samples. Recent self-supervised approaches, such as GLASS, learn normal visual patterns using only defect-free data and have shown strong performance on industrial benchmarks. However, their computational requirements limit the deployment on resource-constrained edge platforms, and even more so within in-sensor processing architectures. This work introduces TinyGLASS, a lightweight adaptation of the GLASS framework designed for real-time edge and in-sensor anomaly detection. The proposed architecture replaces the original WideResNet-50 backbone with a compact ResNet-18 and introduces deployment-based modifications that enable static graph tracing and INT8 quantization. We evaluated the proposed approach on the Sony IMX500 intelligent vision sensor, exploiting the in-sensor processor using the Sony Model Compression Toolkit. In addition to evaluating performance on the MVTec-AD benchmark, we investigate robustness to contaminated training data and introduce a custom industrial dataset, named MMS Dataset, for cross-device evaluation. Experimental results show that TinyGLASS achieves 8.6x parameter compression while maintaining competitive detection performance, reaching 94.2% image-level AUROC on MVTec-AD and operating at 20 FPS within the 8 MB memory constraints of the IMX500 platform. System profiling showcases low power consumption (4.0 mJ per inference), real-time end-to-end latency (20 FPS), and high energy efficiency (470 GMAC/J). Furthermore, the model demonstrates stable performance under moderate levels of training data contamination.