Resolution Matters: An Effective Approach to Anomaly Detection

Abstract

Unsupervised anomaly detection has been profoundly impacted by the advent of large-scale Vision Foundation Models (VFMs). The prevailing paradigm leverages features from a pre-trained encoder, where anomalies manifest as statistical outliers. However, a fundamental challenge persists: industrial defects vary dramatically in scale, making it difficult for a single model to detect them all effectively. This paper introduces a critical insight: a "division of labor" exists between different input resolutions. We find that low-resolution inputs excel at robustly recognizing the presence of anomalies by capturing global context, while high-resolution inputs are essential for refining their segmentation boundaries with precision. To harness this synergy, we propose Multi-Resolution Fusion (MRF), a simple yet powerful training-free strategy. MRF constructs a feature pyramid from the input space by processing an image at multiple resolutions. By fusing features from this pyramid, our method, MRF-AD, effectively combines the recognition capabilities of low-resolution views with the refinement capabilities of high-resolution ones. Extensive experiments show that MRF-AD achieves highly competitive, and in several cases state-of-the-art, results on challenging benchmarks like MVTec AD 2, proving the efficacy of our multi-resolution approach.