Research And Implementation Of Content Based Image Retrieval Methods

With the popularization and development of Internet, the number of image data grows dramatically fast, and how to retrieve image efficiently and quickly becomes an important issue in the field of image's application. In order to manage and retrieve large amounts of images, the CBIR has emerged to be one of the hot research areas in image domain. CBIR extracts visual features as the retrieval features, such as color, texture, shape etc. Generally, an example image is needed when a user uses the CBIR system, which can extract the features of the example image, then compare with the features of other images in the database and show the result to the user. Although many feather-extracting methods and similarity measurements have been raised, they are not mature enough to get good retrieval results. New technology of CBIR researching is required to improve retrieval performance.This dissertation briefly summarizes CBRI system, and researches some key techniques of the image retrieval which specially focuses on feature extraction.A new method for image retrieval based on color histogram of salient points is proposed, which eliminates negative influence of Hu invariant moment image retrieval method based on salient points without considering color information. The experimental result of 1000 images classified in 10 kinds demonstrates its retrieval precision is increased over one time than the image retrieval method of Hu invariant moment based on salient points. The experimental result shows the limitation of the method based on shape feature.Based on the first method be proposed in this dissertation, a new method for image retrieval based on local color distribution of salient points is proposed. An annular color histogram based on local color of salient points is designed, which does not only take advantage of the local color feature, but also takes the space distribution feature of salient points into consideration. The experimental result of the same images database demonstrates its retrieval precision is increased 5.5% comparing with the first method be proposed in this dissertation, and it can increase the retrieval precision 2.5%, comparing with global HSV color histogram image retrieval method.Finally, a prototype image retrieval system MyCBIR is constructed, which realizes the content-based image retrieval. The experiment with the MyCBIR and 1000 images database classified in 10 kinds proves the validity of the proposed methods.