Research On Retrieval Mechanisms For Images And 3D Objects

In recent years, content-based visual information retrieval has become an active research direction in multimedia analysis. The most important problem in this field are how to develop appropriate mechanisms to get satisfactory retrieval results. In this dissertation, we focus on retrieval mechanisms for images and 3D objects. The research is mainly about region-based image retrieval and view-based 3D object retrieval. The contributions of our work include the following parts:(1) we carefully analyse several mainstream feature extraction methods in theory which are belonged to CBIR system, and evaluate the advantages and disadvantages of various methods, then make a series of tests based on search results in the MATLAB platform, and analy the experiments of today's mainstream feature extraction.(2) There are several drawbacks in feature extraction of CBIR system, we proposed the application of Zernike orthogonal moments to describe an image feature, as the higer order moments can be constructed by Zernike moments freely. The image recognition based on this feature is better than other methods, then this method is proved by experiment tests.(3) The defect of CBIR system is a single search method, we proposed hand-drawing as a new search method, which has the advantage of flexible form, close to the high-level semantic retrieval method and so on, with strong application value. (4) Based on Zernike moments, we proposed a measure mechanism of the distance between the 3D object. The distance measure that takes into account the overall properties of each object, but also considered the different matching relations between object views. Experimental results showed that the distance measure in 3D object retrieval is effective. (5) Extended the above algorithm, we proposed a classification of relevance feedback learning algorithm. Accumulate user annotation information in retrieval processes, and extract the image semantic content knowledge adaptively. Then the system incorporates low-lever features, relevance feedback learning, together to retrieve images. Experiments on NTU database show the effectiveness of our methods.