Research And Application Of Image Retrieval Technology Based On Feature Points

With the rapid development of multimedia information technology, the capacity of the databases which store a variety of multimedia resources is being increased; image database is a representative example. Faced with large-capacity image database, users urgently need a rational approach to quickly and accurately locate the images corresponding with the specific content and use these image resources to facilitate the work-related. Therefore, in recent years, content-based image retrieval has become an important research direction in the domain of image processing.The main research work of this thesis is as following:(1) This thesis has done an in-depth study on the commonly used content-based image retrieval methods, and divided the content-based image retrieval methods currently used into three categories: making use of low-level features for searching, combining searching with machine learning methods, and retrieval with relevance feedback. Meanwhile, the basic principles of each category of methods are studied and their advantages and disadvantages are analyzed.(2) An image retrieval method based on feature points is proposed in this thesis, which belongs to the image retrieval methods that use low-level features for searching. and its innovations are as follows: it follows the aesthetics-based principles of composition while dividing the image into blocks, and highlights the importance of the main object in image through the overlapping of adjacent blocks. It extracts both Harris corner points and DOG points in each block, which can express more completely the structure and shape information, comparing with the use of a single kind of feature point. After extracting feature points, it obtains the color and edge information in the neighborhood of feature points by calculating the color moment and the gradient, which, compared with the use of a single visual feature, can describe the characteristics of the neighborhood of feature points more fully and thereby increase the accuracy of the feature point neighborhood matching in the process of image retrieval.(3) Validity of this method is verified through a series of experiments. Experimental results have shown that compared with the single feature point-based and single visual characteristic-based methods, this method can effectively improve the accuracy of image retrieval. In addition, this method is applied in a number of scenes for the experiment, and it achieves good results.Finally, the innovations of the method proposed by this thesis are summarized, and the parts of the method that can be further improved are also discussed.