Product Image Retrieval Based On Shape

With the rapid development of computer networks and multimedia technology, the application areas of images become increasingly wide. The text or keyword based retrieval is becoming hard to meet the requirements in real applications, consequently the technology of content based image retrieval comes with the tide of fashion. As the products in the e-commerce are usually in the form of images, the technology of content-based image retrieval should play an important role in the e-commerce applications. As image shape carries more semantic information than other features, such as color, texture and spatial position relationship, there exist more difficulties in the implementation of shape retrieval. In this thesis, we focused on the shape-based product images retrieval and solved the various problems appeared in the process. The main works are as followings:(1)The key technical process of image retrieval and some edge feature extraction methods were introduced. Firstly, the existing shape descriptors based on contour and region were discussed, then the principles of the two methods and the advantages and disadvantages of these descriptors were analyzed.(2)A product image retrieval system was developed based on the shape context algorithm. This method uses the image contour feature to retrieve. Firstly, a small number of discrete points on the contour were extracted to represent the image's shape. Then a vector set of a sample point to the rest points was obtained as the spatial information. Finally, the feature points of two shapes were matched according to their shape histograms. The product image retrieval system had good retrieval results when tested product images changed on rotation, displacement and scale.(3)A second level retrieval algorithm based on color layout and pyramid histogram was proposed and developed. The first-level was the classification algorithm based on color layout, in which the image was categorized into the product class of the highest probability in the retrieval result. The second-level algorithm was based on pyramid gradient histogram, which tested images with their pyramid gradient histograms in the certain class dataset predicted by the first-level. Both the two levels could meet the real-time requirement. The pyramid gradient histogram were extracted both with local and global features, consequently the second-level retrieval algorithm could retrieval in accuracy and real time.