DiffusionAD: Norm-guided One-step Denoising Diffusion for Anomaly Detection

TL;DR

DiffusionAD presents a diffusion-based anomaly detection pipeline that reinterprets reconstruction as a noise-to-norm process, enabling anomaly-free restorations. It introduces a fast one-step denoising regime and a norm-guided, multi-scale fusion to handle diverse anomaly manifestations, followed by a segmentation module that detects pixel-level anomalies from input versus reconstruction. The approach achieves state-of-the-art results on multiple industrial datasets and maintains real-time inference speeds, addressing both reconstruction quality and practical deployment concerns. Extensive ablations validate architectural choices, denoising strategies, and the anomaly synthesis scheme, underscoring the method's effectiveness and generality for industrial anomaly detection and localization.

Abstract

Anomaly detection has garnered extensive applications in real industrial manufacturing due to its remarkable effectiveness and efficiency. However, previous generative-based models have been limited by suboptimal reconstruction quality, hampering their overall performance. We introduce DiffusionAD, a novel anomaly detection pipeline comprising a reconstruction sub-network and a segmentation sub-network. A fundamental enhancement lies in our reformulation of the reconstruction process using a diffusion model into a noise-to-norm paradigm. Here, the anomalous region loses its distinctive features after being disturbed by Gaussian noise and is subsequently reconstructed into an anomaly-free one. Afterward, the segmentation sub-network predicts pixel-level anomaly scores based on the similarities and discrepancies between the input image and its anomaly-free reconstruction. Additionally, given the substantial decrease in inference speed due to the iterative denoising nature of diffusion models, we revisit the denoising process and introduce a rapid one-step denoising paradigm. This paradigm achieves hundreds of times acceleration while preserving comparable reconstruction quality. Furthermore, considering the diversity in the manifestation of anomalies, we propose a norm-guided paradigm to integrate the benefits of multiple noise scales, enhancing the fidelity of reconstructions. Comprehensive evaluations on four standard and challenging benchmarks reveal that DiffusionAD outperforms current state-of-the-art approaches and achieves comparable inference speed, demonstrating the effectiveness and broad applicability of the proposed pipeline. Code is released at https://github.com/HuiZhang0812/DiffusionAD

Figures (13)

• Figure 1: Anomaly detection and localization examples on MVTec bergmann2019mvtec and VisA zou2022spd. Compared to the previous autoencoder-based approach DRAEM zavrtanik2021draem, our proposed DiffusionAD exhibits superior reconstruction quality consisting of anomaly-free recovery of anomalous regions and fine-grained reconstruction of normal regions. Moreover, DiffusionAD locates the various anomaly regions more accurately.
• Figure 2: Comparison of different algorithms on image AUROC and inference speed. The Y-axis indicates the anomaly detection capability. The X-axis refers to the inference speed. These results are verified on the VisA zou2022spd dataset.
• Figure 3: An overview of the proposed pipeline DiffusionAD. The reconstruction and segmentation sub-networks constitute the entire pipeline. The input image is perturbed by two distinct noise scales. Following, the noise is predicted by an inference within the diffusion model. Finally, the norm-guided one-step denoising paradigm is employed to predict an anomaly-free reconstruction. The segmentation sub-network predicts pixel-wise anomaly scores by comparing commonalities and inconsistencies between the input image and its reconstruction.
• Figure 4: Observation i@. When the noise scale is less than 500, the anomaly-free reconstruction results obtained through one-step denoising are comparable to those obtained through iterative denoising, as indicated by their low mean square error (left) and similar perceptual quality (right).
• Figure 5: Observation ii@. The impact of two different noise scales ($t_s=200$ and $t_b=400$) on the anomaly-free reconstruction for different types of anomalies. Images with green borders represent superior reconstructions.