Bidirectional Condition-Guided Denoising Diffusion Model for Industrial Anomaly Detection

Unsupervised anomaly detection (UAD) is crucial in industrial applications, and condition-guided denoising diffusion models (DDMs) have attracted increasing attention because of their strong reconstruction capabilities. However, existing guidance conditions do not effectively decouple the reconstruction requirements of normal and anomalous regions, making it difficult to simultaneously achieve consistent reconstruction in normal regions and differentiated reconstruction in anomalous regions. To address this challenge, we propose a bidirectional condition-guided DDM (BCGDD), whose core lies in constructing bidirectional guidance conditions to decouple the conflicting reconstruction requirements across different regions. First, we construct a memory bank containing normal patches and their corresponding encodings to provide anomaly-free components. Subsequently, we propose an adaptive fusion strategy (AFS) that adaptively fuses the input patch with its anomaly-free component on a continuous scale to construct bidirectional guidance conditions that simultaneously satisfy the reconstruction requirements of different regions. Meanwhile, we design a hierarchical encoding matching (HEM) strategy that accelerates the retrieval of anomaly-free components through two-level indexing. Experimental results on challenging industrial benchmarks demonstrate that BCGDD achieves state-of-the-art (SOTA) performance, with pixel-level average precision (AP) scores of 79.7% and 59.3% on MVTec AD and KolektorSDD, respectively.