Multi-view patch inference with frozen DINOv3 for industrial anomaly detection

Abstract

Industrial anomaly detection is a critical task in automated visual inspection systems, aiming to identify subtle defects without relying on large-scale anomaly annotations. Recently, pretrained vision transformers such as DINO have demonstrated strong feature representation capabilities and have been widely adopted in this domain. However, most existing methods rely on single-grid patch partitioning, which suffers from boundary information loss and limited local context modeling, resulting in degraded performance on small or boundary-located defects. In this paper, we propose a simple yet effective framework built on a frozen DINOv3 ViT-B/16 backbone for industrial anomaly detection. To address the limitations of conventional patch-based inference, we introduce a Multi-view Patch Inference (MVPI) strategy, which constructs multiple overlapping spatial views of an input image. Each view is encoded independently by DINOv3, and overlapping regions are merged by a deterministic mid-point rule, so that no output token lies on a view boundary. On top of the extracted features, we further propose a Multi-scale Reconstruction Fusion Transformer (MRFT), in which lightweight reconstruction heads model the normal feature distribution of each backbone layer via cosine similarity-based reconstruction learning. The anomaly score is computed from reconstruction residuals, enabling effective discrimination between normal and defective regions. Extensive experiments on three benchmark datasets, including MVTec AD, BTAD, and UltraChip, show that the proposed method achieves state-of-the-art or competitive performance in both image-level and pixel-level anomaly detection, while requiring no memory bank and only moderate GPU memory at inference. The code will be released at https://github.com/chaoqunWang666.