Content-Based Retrieval of Surface Defect Images with PicSOM

In this paper a content-based image retrieval (CBIR) system called PicSOM is applied to a defect image database containing 2004 images from a real web inspection system. The main features of PicSOM are efficient indexing based on self-organizing maps and adaptive querying using relevance feedback. Six feature descriptors from the MPEG-7 standard and an additional shape descriptor developed for surface defect images are used in the experiments. The classification performance of the descriptors is evaluated using K-Nearest Neighbor (KNN) leaveone-out cross-validation and PicSOM’s built-in CBIR analysis system. The KNN results show good performance from three MPEG-7 descriptors and our shape descriptor. The CBIR results using these descriptors show that PicSOM’s SOM-based indexing engine together with its relevance feedback mechanism yields efficient and accurate retrieval of similar defect images from our database.