Research On Relevance Technology Of Region-based Image Retrieval

With the rapid development and wide application of computer and multimedia technology, the digital image increases at a great rate. How to organize, analyze and retrieve it well gradually becomes a hot subject.This thesis focuses on relevant technology of region-based image retrieval system, which includes the extraction of underlying visual features for images, region-based image segmentation, similarity computation, relevant feedback and so on.The technology of extracting underlying visual features for images is one of the crucial processes to CBIR, which points a situation that at present the most image segmentation algorithms operates too slowly and lacks the real-time performance, thus proposing a new features scheme that divides images into blocks based on L*a*b* color space, which could reduce the sample data that is pending so as to effectively saves the time wasted by the image region segmentation.As for the region-based image segmentation, there is a new and fast algorithm. This method, based on extracting underlying visual features for images, firstly applies Mean-shift clustering algorithm to complete the cluster so as to obtain the number of cluster and the initial centre; then by using the modified K-means algorithm it achieves the fast division for images. The result of experiment shows that by using the new method not only does the rate of segmentation increase, but the outcome is also stable with avoiding the excessive division.For the image similarity computation, the thesis introduces an approach based on integrated region, which takes all the regions of the segmented image into account. The experiment shows that this measure has better performance. On the basis of similarity computation it presents an idea about the constrained random browse, which lets the users acquire the image database rapidly and allow them select an image to be the inquiry example.For the relevant feedback, it introduces a SVM-based and feedback-log-file-based image query method, which aims to make the retrieval results become more useful to the users by using the interactive and feedback data.Based on above researches, this thesis designs and implements the region-based image retrieval system, called IERBIR, which establishes the feature data bank with SQL to store image features. The experiment proves that its retrievable performance is fine.