Method Research Of Content-based Image Retrieval By Texture And Shape

With the development of the eras,the text and multimedia information in the Internet is increasing at an unprecedented rate.Under such a background,content-based image retrival is one of the most important areas of computer vision.two topics have been studied and developed: 1.the extraction of image color,texture,shape and other features;2.perfect effective retrieval strategy.On this basis,this paper integrates two algorithms of texture features extraction:Gabor color moments and Bo W of SIFT.Combined with shape context algorithm for shape features extraction,a content-based image retrieval system is constructed.In terms of feature extraction,this paper discusses all kinds of image color spaces and compares the advantages and disadvantages of various color spaces,while studies the observation theory of image localization based on Gabor wavelet transform in time domain and frequency domain.The multi-scale and multi-directional Gabor filter is set up and applied to the convolution operation of the image to obtain the multi-scale texture features of the image.At the same time,four important implementary steps of SIFT feature extraction are discussed.Combined with K-Means algorithm,a number of SIFT feature vectors are transformed into a visual word feature vector,which is mainly composed of clustering centers.The extracted texture features are linearly fused and normalized to generate a high dimensional fusion feature vector.Canny operator and adaptive double threshold algorithm are selected as the edge detection operators in this paper,and shape context feature is extracted as a shape feature extraction scheme described by distance and location relation between point sets.In terms of retrieval strategy,for high dimensional feature fusion,this paper proposed an retrieval algorithm to judge confidence of picture which will be returned,by cooperating vote mechanism of multi-classification support vector machine with distance measures for feature vector.transform function of support vector machine from sample classification into sample confidential judgement.Effectively enhance the retrieval accuracy in the case of classification error.While the shape context features are matched by the Hungarian algorithm to return the appropriate retrieved images.Finally,the article establish a Content-based Image Retrieval System,validate robustness of algorithm at VOC-2007?CIFAR-10 dataset.at the meantime validate effectiveness of algorithm is carried out on Corel-1K image dataset and self-built image dataset.Retrieval result on both dataset is better than system built by single feature and traditional distance measures algorithm.