Self-supervised Defect Detection and Localization Based on Heatmap
Pseudo Anomalies (HPA)
Abstract
Anomaly detection is widely used in manufacturing and medical imaging.
We propose a self-supervised defect detection method based on
multi-scale feature fusion, which can effectively improve the detection
and localization accuracy. The method of pseudo-defect construction was
used to enhance the training data. To make the pseudo-defects more
realistic, the extreme point of feature heatmap was used as the anchor
point of the defect area, and the defect image was fused with the
original image to construct the pseudo-defect. A multi-scale feature
fusion network was proposed that utilizes the self-attention mechanism
and the interaction between multi-scale features to extract semantic
features containing rich contextual information to improve detection and
localization accuracy further. The proposed method achieved competitive
experimental results on both the MVTec AD and Chest X-ray datasets.
Compared with other pseudo-defect simulation methods, the heatmap-based
pseudo-defect construction method improves by at least 2%. It achieves
comparable results with other state-of-the-art defect detection methods.