Research On Image Retrieval Method Based On Multi-Instance Learning

In recent years, as the development of multimedia and Internet, more and more attention have been paid to image information. How to get the required information from a mass of database efficiently, quickly and accurately became a most important problem. Therefore the technology of CBIR(Content Based Image Retrieval) appeared and became a hot spot in the field of multimedia search.In this thesis, related research on CBIR and the basic theory of multi-instance learning were introduced, and the current field of multi-instance learning application were analysed. A multi-instance learning based on CBIR approach had been advanced in order to deal with the problem of inherent ambiguity of images. First of all, the whole image was regarded as a multi-instance bag. Secondly the image was partitioned into a number of regions by adaptive k-means image segmentation algorithm. Thirdly EM-DD(expectation maximization diverse density) algorithm was used for multi-instance learning, and the images as learning results were ranked according to their similarity. Finally, corresponding feedbacks were implemented on the basis of the positive bags and negative bags which user had chosen, and satisfied user with the ultimate result.