Clothing Retrieval Based On More Characteristic Index And The Process Of The Geometric Validation

With the rapid development of the electronic commerce, the emerging of on-line shopping mode has gradually replaced the mode of traditional shop shopping. Especially in the shopping of clothing, on-line shopping has a greater advantage than traditional shopping, requiring a considerable amount of energy to run back and forth to the store. We just by clicking the mouse, and never leave home can browse all kinds of clothing on the Internet. All we have to do is a simple click, then we can view various kinds of clothing and buy what we like within home. But unfortunately, the apparel search of most on-line shopping websites is based on product labels as the mainly form of information retrieval. Because there are great subjective differences in linguistic description and worse still, it is impossible for people to use the accurate words or labels to describe the clothes they like when they come across beautiful clothes on street, TV or websites. Thus, it is hardly for people to search for the same style of clothing via websites by using some description words or key labels. Therefore, it is of vital importance to design a system by utilizing objective images to retrieval similar images. In this system, it is possible for people to get the similar clothes accurately and precisely by inputting the original clothing images.This paper proposed a method based on the multiple features indexes and the process of post verification. The accuracy and performance of this method have been verified by a variety of experiments. The contributions of this paper consisted of four main aspects.(1)The first aspect:Take the advantage of SMQT features and SNOW classifier to perform face detection in the process of preprocessing image. The accurate position of human face could be identified by using the face detection method when a complete image has been input. Then, this paper got the clothing areas in this image by utilizing the Grab Cut algorithm to split the image after getting the accurate positions of clothing. Just the clothing areas would be saved after the process of preprocessing.(2) This paper combined the color descriptor(CN) with the traditional BOW framework to enhance the matching of color features. In the meanwhile, this paper has adopted SIFT feature and color feature into the framework of multidimensional index for feature matching.(3) As for the reducing the discrimination ability of the local area and losing the geometric relationship between features after using BOW model to quantify the visual words, this paper put forward geometric post verification based on feature sizes to test wrong matching features. This method was able to eliminate the wrong matching points and remark candidate images by analyzing the relationships of features.(4) Tested the performance of the method proposed in this paper via a great many of experiments. Then, this paper compared the accuracy of each improvement of the retrieval framework. Finally, the experiments’ results showed that the method proposed in this paper could achieve a desirable and good performance even being used to process massive images in database.