EdgeDiff: Leveraging Edge Maps for Anomaly Detection with Diffusion Models

Anomaly detection is a challenging but significant task in many applications. Existing methods based on reconstruction typically use the original images as inputs. However, these methods are usually limited by inaccurate reconstruction. To overcome this weakness, we propose an anomaly detection method named EdgeDiff based on the diffusion model. It achieves the reconstruction of the anomalous image by mapping the edge map of the image back to the original image. By utilizing the high-frequency information reflected by the edge maps, anomaly regions can be reconstructed more accurately while preserving the normal regions. We compare the original and reconstructed images in the feature space, and investigate how to utilize Segment Anything (SAM) to further improve the anomaly detection performance. Extensive experiments on the MVTec-AD benchmark have shown the effectiveness of our method.