LaRHP: latent-aware reconstruction via hypersphere projection for industrial image anomaly detection.

Industrial image anomaly detection identifies defects by distinguishing abnormal from normal patterns. Due to limited abnormal samples, semi-supervised and unsupervised approaches are more practical. Among these, reconstruction-based methods face a trade-off: larger latent dimensions reconstruct anomalies too well, while smaller latent spaces degrade normal reconstructions, making it difficult to keep false negatives and positives low. To address this, recent reconstruction-based methods integrate feature-embedding constraints into their training objectives. Feature-embedding approaches effectively model normal distributions by defining decision boundaries in latent space but require extra processing for precise anomaly localization. Consequently, hybrid methods combine reconstruction and embedding, yet still struggle with this trade-off. We propose LaRHP, a latent-aware reconstruction method enhanced by a hypersphere projection mechanism that integrates strengths of both approaches to overcome the balancing challenge. LaRHP's innovation lies in its projection mechanism, applicable across different reconstruction-based architectures. LaRHP extracts multi-scale features with a pre-trained network and employs an autoencoder with dual objectives: accurate reconstruction and compact latent representations constrained to position-specific hyperspheres. This alleviates distance concentration issues in high-dimensional spaces, improving anomaly discrimination. During inference, anomalies outside the learned hypersphere are projected onto its surface, reconstructing their closest normal counterparts and amplifying reconstruction errors in anomalous regions without affecting normal parts. Evaluations on MVTecAD show LaRHP significantly improves defect localization (pixel-level AUC-ROC 0.97, + 1.8%; region-level AUC-PRO 0.92, + 1.2%) and achieves competitive classification performance (mean AUC-ROC 0.953). Generalization analysis confirms broad applicability across reconstruction architectures and datasets, with latent space visualizations revealing meaningful representations and superior detection performance.