Optimized Visual Anomaly Detection Models for Tiny Machine Learning

This thesis explores the field of visual anomaly detection within the context of Tiny Machine Learning (TinyML), focusing on the optimization of existing models for deployment on resourceconstrained edge devices. We begin by benchmarking several state-of-the-art unsupervisedVAD models,such as PaDiM and PatchCore,on popular datasets like MVTec, ViSA, and Real-IAD. These benchmarks provide a performance baseline for comparison. We then explore how these models behave when trained on noisy or contaminated data. Finally, we apply specific optimizations to improve the models’ performance on edge devices. These include simplifying distance calculations for PaDiM and applying quantization techniques to reduce memory usage in PatchCore. Our results show that it is possible to significantly improve the efficiency of these models without heavily impacting the performance of the original models, making them more suitable for real-world deployment on low-resource devices.