BoRAD: Bootstrap your Own Representations for Multi-class Anomaly Detection

A new training framework called BoRAD aims to improve multi-class anomaly detection for industrial inspection by treating the problem as a representation-capacity allocation challenge, according to a preprint posted to arXiv on June 12, 2026 [1][2]. The framework, detailed in a paper submitted to the computer-vision section of the open-access repository, addresses a core difficulty in scaling anomaly detection from single-category models to a unified, one-for-all setting [1][2]. Reconstruction-based methods are widely used for spotting defects in manufacturing, but a single model must learn to reproduce a wide range of normal product appearances without inadvertently copying anomalies. The authors identify two coupled failure modes that arise: an identical shortcut, where anomalies pass through the reconstruction path, and mis-reconstruction, where normal categories are confused with one another [2]. BoRAD, which stands for Bootstrap your Own Representations for Multi-class Anomaly Detection, is a label-free framework that uses a shared learnable prototype bank during training [2]. It imposes two complementary regularizers. Spatial prototype alignment contracts local within-prototype variation to suppress anomaly copying, while prototype-relative global alignment preserves between-prototype structure and improves sensitivity to abnormal angular deviations [2]. The prototype bank and prediction heads are discarded at inference time, leaving a standard teacher-student feature discrepancy pass that requires no class labels, negative pairs, memory retrieval, or prototype lookup [2]. The paper reports competitive results on three benchmark datasets. BoRAD achieves a mean anomaly detection score of 86.2% on MVTec AD, 80.7% on VisA, and 73.1% on Real-IAD [1][2]. Diagnostic analyses accompanying the results show reduced anomaly leakage, improved separability among normal categories, and stronger separation between anomaly and normal scores [2]. The preprint appeared on arXiv, an open-access repository that hosts e-prints in fields including computer science and electrical engineering and has grown to a submission rate of about 24,000 articles per month as of late 2024 [6]. The paper’s abstract page also features arXivLabs integrations, a framework launched in 2020 that allows community collaborators to build experimental tools such as citation explorers and code finders directly on the site [4][5].