Research On Multiple Features Fusion Hierarchical Image Retrieval And Implementation

Content-based image retrieval is an important research direction of multimedia information retrieval. It is based on the idea of "In Fig Search Map" and widely used in many fields, such as product image retrieval in e-commerce. With the popularity of the network, digital images have increased significantly, and become main information sources of multimedia. Various images have pervaded our daily life. How to analyze and retrieve the image quickly and accurately is very valuable for research.This thesis makes a deeply research on the content-based image retrieval technology. Based on learning the relevant theoretical knowledge and algorithm, we use a hierarchical image retrieval method based on multi-feature fusion. In this method, the first layer is image filtering. A multiple category classification method based on sparse to filtrate image is presented. The second layer is accurate image retrieval. The multi-feature fusion method is adopted into accurate retrieval images. In this thesis, the multiple feature fusion of hierarchical image retrieval method is applied into product retrieval. The specific work and innovation points are as follows.(1) In image filtering layer based on sparse classification, this thesis proposes a multi-class classification based on sparse approach. The traditional sparse classification method classifies image on the basis of the residual, and the image is classified as a certain kind. It easily leads to classification error and can’t achieve the goal of image filtering. Therefore, a multi-class classification approach combined with the characteristics of sparse representation is proposed, which uses residual and the ratio of sparse coefficient to classify images into the first N similar classes to reduce classification error. In this way, we achieve the purpose of image filtering and form a set of candidate dataset for accurate image retrieval.(2) In accurate image retrieval layer, this thesis uses a retrieval method of multiple features weighted fusion. Firstly, we extract the local feature of SIFT. The histogram feature of the image is obtained by the Bag-of-Words model. The Dirichlet Fisher kernel is introduced into the histogram-based feature to obtain SiftBowDirfk feature with strong discrimination and improve the performance of retrieval. In addition, in terms of single feature cannot accurately describe the image characteristics, we use multiple feature weighted fusion method to represent the image. Thus the representation ability of the image characteristics is enhanced and the image retrieval precision rate and recall rate is improved.(3) Combined with hierarchical image retrieval technology, this thesis proposes a sparse multiple features fusion of hierarchical image retrieval method. The number of matching features in retrieval is reduced via filtering images with sparse multi-classification. The retrieval performance is improved by image retrieval with multi-feature weighted fusion method.(4) The multiple features fusion of hierarchical image retrieval method is applied to product retrieval, thus a hierarchical product retrieval system is designed and implemented.