Reseach On Improved SURF Descriptor And Its Application In Clothes Image Retrieval

The rapid development of electronic commerce leads to the tremendous increase of apparel goods data. It becomes more and more difficult for users to choose what they want from huge amounts of apparel goods. Most of the information about apparel goods is displayed with clothes images. This leads to the generation of the retrieval system of clothes images.Texture feature is a kind of important visual characteristics. It is widespread but difficult to describe in the image. It contains the surface information as well as its relationship with the surrounding environment. The macrostructure and microstructure of the image are both combined. This thesis focus on the local texture descriptor aimed at clothes images and its application in clothes retrieval system. The research is mainly about the Fast Hessian feature point detection algorithm, an improved local descriptor, the matching strategy, similarity computing, and their application of clothes retrieval system. The main work of this thesis is divided into four parts:(1) Research and realizes the feature point detection algorithm called Fast Hessian. We get the coordinate of the interest points and in which scale they are detected successfully. And these interest points are verified to be effective in texture recognition of clothes images.(2) To improve the instability of rotational invariant taken from the using of dominant orientation in SURF descriptor, we propose an improved local descriptor which is invariant to rotation itself using intensity order pooling. Next, we put forward a division strategy using concentric rings to keep part of the loss spatial information brought from grouping by intensity. Finally, we use accumulation to construct the descriptor vector through a contrast experiment.(3) The texture of clothes is always repeated, so we propose the multi-neighours strategy to match the descriptors and define the computing of image similarity and average similarity to measure the similarity degree of texture between images. The similarity value is verified by experiment that it can effectively measure the similarity of image texture in most cases. We also set a contrast experiment between our local descriptor and SURF to compute the similarity value of rotationed images with repeated patterns. The result shows our descriptor performs better and is more stable.(4) A clothes retrieval system based on texture is proposed in the final part. The main architecture of the system design and the structure and function of each module are analyzed. Finally, the clothes retrieval system is implemented. The accuracy of texture recognition and the effect of clothes retrieval are verified through tests.