Image Retrieval Based On Partition For Interest Points And Text-based IR

Recent image retrieval systems make direct use of the native content of the images. The low-level visual features, such as color, texture and shape used by most existing content-based systems, however, have little correspondence to the semantics of the image and are difficult to be understood by users. The user of a retrieval system, on the other hand, most likely associates image similarity with semantics. Thus, high-level semantic concepts play a large role in the way we perceive images and measure their similarity.In this paper, a new approach of image retrieval combined interest points and text-based IR is proposed. Firstly, after the current methods of the interest points extraction are discussed, a new image representation method is proposed, which uses a visual attention model to extract focuses of attention in an image and makes them as interest points. Then, keyblocks are generated by means of partitioning for interest points and are described by color feature. This image representation method has many advantages. It can reduce redundancy, and can avoid the difficulty of image segmentation and object-oriented descriptions semantically. At last, clustering the eigenvectors of keyblocks to generate the codebook, so the images in image database can be represented by one dimension data, which is similar to the text data structure. At this time, text-based IR technology can be used to content-based image retrieval. Through Boolean Model of inverted file and Histogram Model, a new index is established aimed at syntactically representation of object content of image. So the"semantics gap"between content-based system and the concept-based user is simply connected.