An Image Retrieval Method Based On Gaussian Mixture Model

CBIR(Content Based Image Retrieval)is one of the popular topics in current image retrieval. How to describe image characters efficiently and accurately is a core problem in CBIR. Based on content, this thesis discusses an extraction method and designs an image character extraction system by using extracting color characters of the chromatic image. The research contents and the originalities are as follows:This thesis has a deep research on the application of the color character in the image retrieval and discusses the key problem: color description and extraction method of color character, when it comes to image retrieval by using color characters. After discussing advantages and disadvantages of SQ histogram generation method and VQ histogram generation method in HSV space, this thesis puts forward an improved VQ histogram generation method, which is a kind of vector quantification histogram generation method based on Gaussian mixture model, and makes comparisons about the performances among SQ, VQ and VQ methods when they are used to retrieve image. At last, in the retrieving process, we apply different histogram distance, Euclidean distance,Intersection distance,Quadratic distance to these generation methods。To prove the effect of the method, this thesis firstly sets up a demonstrate system, then generates histograms of each image in the database using SQ,VQ,GMVQ methods respectively, and makes the image retrieval by using the histograms as image characters. Meanwhile, during the retrieval process, it uses different matching algorithm based on the similarity of the histogram. Finally, This thesis has a computing and analysis on the recall and precision of the retrieval method based on GMVQ and makes a comparison between the histograms generated from SQ,VQ methods. Experiments prove the method put forward in this thesis is better than the other methods as to the retrieving performance.