Researches Of Multi-Instance Learning On Content Based Image Retrieval

Nowadays, with the development of multimedia technology and Internet technology, image information gets more and more people’s attention. How to find effectively and accurately the interested image from a large image database, has become one of important problems which need to be solved urgently. Content-based image retrieval emerges at this time, and becomes a hot issue in the field of multimedia information processing. Content-based image retrieval has made great progress, but the use of color histogram or other global feature make retrieval image content is ambiguity. An image contains typically more content cannot adequately describe more content only by a single global feature, and also cannot accurately express interested content. Multi-instance Learning method can better solve the ambiguity.By analyzing the existing image retrieval methods based on Multi-instance Learning, it is not difficult to see that the main idea is that transform an image into an instance bag. And then to learn the given search results through the Multi-instance Learning algorithm. So the generating method of Multi-instance image bag determines the whole learning process quality. The dissertation put forward two kinds of generating methods of image bag.The work contents in the dissertation as follows:1. The dissertation introduces research status of content-based image retrieval at home and abroad. It introduces theory knowledge of multi-instance learning. It analyzes the corresponding relationship of image retrieval and Multi-instance Learning, and also analyzes the existing image retrieval method based on Multi-instance Learning.2. By analyzing the existing image retrieval methods based on Multi-instance Learning, we put forward two kinds of generating methods of image bag based Scale-Invariant Feature and Bag-of-Words. We give the generating process of image bag and the corresponding image retrieval process based on Multi-instance Learning. We make comparative experiments with image database. It shows these two kinds of method have better precision, and also illustrates these two kinds of generating methods are effective.3. When learning in image retrieval process based on Multi-instance Learning, the attribute dimensions of instance have a great influence on learning time efficiency. With the development of research of multi-instance learning, it must deal with large amount and high-dimensional data. So we put forward Multi-side Multi-Instance (MSMI) Algorithm. By choosing reasonably attribute, the process of Multi-instance Learning become sample. And the experiments on benchmark data sets show that it is effective.4. The dissertation gives a summarization and presents the work we will do in the future.