A Statistical Approach to Industrial Anomaly Detection

In this paper, we present a statistical approach to anomaly detection and monitoring through image analysis, and its application in non-destructive evaluation. A non-parametric statistical model is created by Parzen window density estimation at each pixel location, based on the observations of a number of defect-free images and the derived low level features. A test image is compared against the learned statistics and pixels not fitting the model are called out as defects, which include material changes and manufacturing defects. Image normalization and registration are carried out to factor out the image-to-image variations of appearance change and spatial misalignment, and anomaly detection on multiple views are combined for overall detection. The assisted defection recognition and monitoring can dramatically increase screening efficiency and consistency, thus improving quality control.