LoGics: Anomaly detection of key components of high-speed rail via zero-shot learning

Abstract

In this paper, We present LoGics (Local Gaussian Distribution with Interaction Scores), a zero-shot anomaly detection framework for high-speed rail component inspection. The method combines three technical enhancements: local Gaussian distribution modeling with sample relationship analysis to enhance anomaly discrimination in training-free settings, feature space normalization addressing cross-domain discrepancies in industrial feature representations and distribution-guided threshold computation that automatically selects optimal strategies based on data characteristics. Experimental results on MVTec and VisA industrial benchmarks demonstrate the framework's effectiveness, achieving 97.9% image-level AUROC. This performance approaches supervised methods while requiring no annotated training samples, suggesting practical value for infrastructure inspection tasks where defect examples are unavailable.