Image Retrieval Method Based On Regions Of Interest

With the development of digital media technology, more and more digital images are produced. How to organize, store, represent,query and retrieve these mages is an urgent problem. The Content-Based Image Retrieval (CBIR) provides a helpful method for this goal. There are three key issues in CBIR: 1) how to represent the content of an image, 2) how to measure the similarity between two images based on these representations and 3) how to make the retrieval system learn human beings'subjective manners in image understanding and comparison.This thesis discusses the first two key issues of CBIR, focusing on image content representation and retrieval methods based on regions-of-interest (ROI).First, the current four image retrieval methods are compared. They are global feature search, color layout search, region-based search and the newly appeared search method based on regions-of-interest. Then, after the current ROI based methods are discussed, a new image representation method is proposed. This method uses a visual attention model to extract focuses of attention in an image and makes them as interest points. Regions-of-interest are the result of interest points clustering. The image is thus represented by a set of regions-of- interest, which are more perceptually important than other parts of the image.To compare the similarity of two images, an image comparison method based on regions-of-interest was proposed in this thesis. This method chooses the most similar regions from the images as matched regions, and compare the visual similarity of the images by content similarity and integrity. The layout similarity, which is invariant to image rotation and flip, is also compared according to the position relationship of matched regions.Finally, we build a prototype system to test the performance of the proposed method.