A Multi - Sample Learning Algorithm For Image Retrieval

With the development of the Internet, and the availability of image capturing devices such as digital cameras, the size of digital image collection is increasing rapidly.As people are increasingly exposed to a lot of image information, effec-tively organize, manage and retrieve large-scale image database has become a problem urgently needed to address. Content based image retrieval has become a hot issue for the field of multimedia search. This paper introduces the research of content based image retrieval and the basic theory of multiple instance learn-ing, analyzes the current applications of multiple instance learning. We propose a merge based multiple instance learning algorithm in the framework of con-tent based image retrieval. The algorithm has the following advantages. Firstly, it solves the shortcoming of depending on unified parameters by making some adaptive improvements of image auto segmentation. Secondly, in the support of multiple instance learning, it takes into account the mutual relation between instances, and solves the drawback of common multiple instance learning algo-rithms which ignore the relationship between instances as they assume that the instances are independent and identically distributed. Thirdly, it upgrades the difference between classes by merging cluster centers, hence increases the retrieval precision. The experiments on the image dataset Corel10000show:the recog-nition rate of our algorithm has reached70%in any20classes. In20selected classes of Corel10000, the rate has reached87.45%. The recognition rate of100classes has reached45.96%. And the TOP3recognition rate has reached66%. Maybe the TOP1recognition rate of our algorithm has not achieved the level of practical applications in large classes scale of, but we still think it is worth mentioning cause it may has its applications in initially image retrieval.