Research On Method Of Integrated Multi-Features Based Image Retrieval

With the pervasion of hardware in image acquiring and pictures storage and the rapid advancement of internet, information from images is playing critical role in lots of aspects in social life. The traditional text-labeled image retrieval can not meet the precise indexing requirements. Content based image retrieval (CBIR) has became the main stream of image retrieval gradually. How to get the wanted images swiftly and effectively in permitted scopes has became the focus of the multi-media researchs.Integrated multi-features image retrieval has became the hot point of CBIR at present. This paper proposes three IMFIR algorithms which are based on non-linear fuzzy histogram, color-texture features and ROI combined with SIFT respectively. The experiments testify the conclusions that the IMFIR algorithms proposed can obtain better performance that the algorithms based on unitary features.So, the contents are as follows:1. This paper summarizes the existing state of the art CBIR techniques such as the construction of CBIR system, features extraction, similarity measurement, the evaluation of retrieval performance, the index techniques of high-dimensional data and relevance feedback. Then, several features of image in color, texture and shape are analyzed and accomplished in this paper.2. After studying the chromatic histograms and gradient vectors histograms, non-linear fuzzy color histogram and non-linear fuzzy gradient vectors histogram, which stand on that the human visual cognition is nonlinear, are created. The histograms combined with some principles of relevance feedback, which are used to optimize the selection of combinatorics of features group, are used to construct the IMFIR algorithm based non-linear fuzzy histograms.3. After the integrated analysis of color and texture distribution, the multi-texton histograms, which are extracted from the original images quantized evenly in RGB chromatic space, are used to measure the texture distributions features; Then, the images are converted into CIE L*a*b* color space and then quantized equably to calculate color coherence vectors as another group of image features. Based on words weighting of text indexing and some principles of relevance feedback, the method of experimental parameters adjusting is used to optimize the weighting value of the two groups of features to composite the IMFIR algorithms. 4. After the analysis of ROI and SIFT, an IMFIR algorithm is proposed. First, in entropy images of original images ROI are extracted. Then the values of Zernike moments are calculated to reduce the size of fist-retrieval images set. At last, SIFT is used to do further matching or retrieval to obtain the ultimate results set. The method is fit for the narrow filed and major image indexing. In the algorithm, lots of parameters are needed to specify, so the relevance feedback is used to learn and adjust the parameters to promote the quality of image retrieval.Images retrieval experiments testify the conclusions that the integrated multi-features could mark the features of image completely and objectively and the relevance feedback could eliminate the gap between image features and visual perception effectively. So, they are very important to construct practical image retrieval system.