Research Summary

Abstract:

Unsupervised visual anomaly detection in industrial and clinical edge environments faces a critical dual challenge: the absolute scarcity of anomalous training samples and the prohibitive deployment costs of high-power GPU infrastructures. To bridge this gap, this paper presents a comprehensive evaluation and deployment framework that cross-analyzes two major deep learning paradigms—generative reconstructive models (Autoencoder, VAE, GANomaly, and Deep SVDD) and non-parametric feature-based descriptors (PaDiM, Patch-Based, and PatchCore)—optimized exclusively for resource-constrained CPU environments. Rather than relying on expensive hardware scaling, our methodology leverages transfer learning via a frozen ResNet34 backbone, combined with geometric Coreset compression and high-performance vector indexing through the FAISS CPU library. Extensive cross-evaluation on the MVTec AD, Industrial Rock, and Chest X-Ray datasets reveals a stark architectural polarization. While generative architectures struggle with semantic representations—manifesting a severe performance degradation on complex clinical textures with a Variational Autoencoder (VAE) AUROC dropping to 0.3301 due to the Blurry Reconstruction Syndrome—the optimized non-parametric feature models achieve a near-perfect AUROC of 0.9991. Crucially, through rigorous software engineering, our unified ecosystem reduces inference latency to an ultra-fast sub-millisecond threshold (< 0.8 ms per frame) on conventional, low-power desktop hardware. Finally, this work formalizes the theoretical boundaries of these paradigms by contrasting the Blurry Reconstruction Syndrome against the metrological paradox of the Gray Zone, where global scalar anomaly score reductions produce marginal false negatives despite flawless spatial pixel-level localization via heatmaps. This benchmark establishes a new blueprint for eco-efficient, highly sustainable One-Class AI deployment at the industrial edge.