Learning for anomaly detection in industrial vision

Details

Supervisors

De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

This thesis focus on detecting defect on images. Three steps are necessary in order to achieve this aim. First two Convolutional Neural Networks are trained over images that have been made artificially noisy. Structured noises like scratch, stain or unstructured noise (Gaussian noise for example) are added on clean images. Secondly CNN try to find a mapping, during the learning phase, between the artificially noisy images and their clean versions. In other words the CNN learns to remove the artificial noise from images. After this training phase a test dataset which contains a mix of real noisy images and clean images is given as input to these CNNs. If the input image is noisy the CNN should output its clean version, and if the input image is clean the CNN should output the exactly same image. Thirdly we compare each input with its corresponding output. If the input and the output are the same the input image shouldn't be noisy, and will be labeled as "clean". But if the input and output are different the input image should contain noise/defects, and will be labeled as "noisy".