Research On Image Analysis For Product Image Search

The rapid development of electronic commerce has changed people's traditional shopping habits.However,the existing search technologies that using classification search or keyword search,have the problems of large amount of search results and low accuracy.This thesis based on the image search technology of commodity,focuses on the core issues of image feature extraction and matching problem.According to problem that traditional feature descriptor SIFT(Scale Invariant Feature Transform)can't make effective matching in case of large affine transformation and perspective transformation,the main work of this thesis is proposing a new image feature descriptor-Affine invariant feature descriptor with multi-angle of views.This descriptor,firstly,generates a set of simulation view image sequence based on the original image;secondly,detects the visual characteristics of the image sequence;finally,uses RANSAC(Random Sample Consensus)algorithm to map these visual characteristics to the original image.All the visual characteristics constitute the feature points of the original imageIn this thesis,the method of constructing the visual dictionary and feature quantization in the traditional BoW(Bag-of-Words)is improved.According to the problem that the K-Means clustering algorithm can't have determined results by random selection of initial clustering center,this thesis proposes Density-sensitive method to determine the initial clustering center.In view of the problem caused by HQ(Hard Quantization)used in visual feature quantization,this thesis proposes SQ(Soft Quantization)to quantify the visual features.Construct commodity image database by the images from taotaosou,zuimeiso,vipshop and so on.Test the method proposed in this thesis,and compare the results with these methods of using HOG,conventional SIFT,and traditional Bo W.It all shows that the proposed image feature extraction algorithm and the improved BoW algorithm is much better.