Real-time pixel based multiscaleanomaly detection inmultivariate images

Sammanfattning: This thesis presents a method using machine vision for surface quality control in an industrial manufacturing process. The method is based on the anomaly detection due to challenging, large variety and low occurrence rate of possible defects that have to be identified. The control is performed on a surface of the inspected product using multiple light sources for better illumination which provides more information about the exterior conditions. The method itself handles the images taken from the different light sources and computes a statistical model of the obtained multivariate image at multiple scales in order to comprehend signaling the possible defects. For the final classification, a pixel based probability map is used to pinpoint possible defects. The method has been implemented in Matlab and tested on a sample set provided from a Swedish furniture factory. Experimental results show interesting properties as well as an evaluation of the proposed method.