AVID-DR: Active-View Selection and Cross-Scale Instance-Aware Anomaly Detection with Diffusion Reconstruction

Abstract

Multi-view visual analysis is often limited by redundant observations, scale-varying anomalies and incomplete image evidence. This paper proposes AVID-DR, a unified pattern-recognition framework that integrates uncertainty-guided active-view selection, confidence-weighted multi-view fusion, cross-scale instance-aware anomaly detection and structure-texture conditioned diffusion reconstruction. The key coupling is a closed uncertainty loop: detector uncertainty, reconstruction variance and fusion confidence are calibrated into a common decision signal that determines whether an additional candidate view should be processed. Candidate views are ranked using low-cost preview cues rather than full downstream inference, and the selected evidence is fused before anomaly localization and conservative reconstruction. The detector combines hierarchical features with prototype-family normal modeling, while the latent diffusion module reconstructs only anomaly-supported regions under structure and appearance constraints. Experiments use structured image degradation with document-image data as the primary empirical setting and separately evaluate localization, reconstruction and view-efficiency. The results are reported over repeated seeded runs; generalization beyond the evaluated structured-image domain is left for future study.