Image Retrieval Technique Based On Multiple Examples

With the development of information technology and internet technology in recent years, the number of images grows explosively. We pay more and more attention on image information. How to find the interest image from a large database effectively and accurately has become one of the hot issues, which needs to be solved urgently. Content-based image retrieval is becoming the focus of attention in academia and industry.Content-based image retrieval has made great progress. But it usually measures the overall visual similarity between images. While the users often judge the similarity of images based on "semantic similarity" rather than "visual similarity". It is difficult for traditional content-based image retrieval techniques to obtain satisfactory retrieval results. In order to establish a bridge between the low-level visual features of images and the high-level semantic, some multi-instance learning methods have been proposed in content-based image retrieval. Regard each image as a bag, the low-level visual feature vectors of the segmented regions as instances, and an image is labeled as a positive bag if the user regards it as semantically relevant. Then the question of classification based on the semantic of image will be turned to multiple instance learning problems.The work contents in the dissertation are as follows:(1)The dissertation introduces research status of image retrieval method at home and abroad. It introduces the multi-instance learning theory briefly and analyzes the existing image retrieval methods based on multi-instance learning.(2) The dissertation introduces the background of multi-instance learning and its basic theory. The differences between multi-instance learning and traditional machine learning are analyzed. Then we introduce the classification algorithm about multi-instance learning and its application briefly. In the meantime, we review the development of image retrieval. Some of the characteristics of the image retrieval were analyzed briefly. The advantages of multi-instance learning combined with content based image retrieval are described. (3) The dissertation analyses of the general process of image retrieval based on multi-instance learning. Image segmentation method based on k-means clustering was introduced and images’ characteristics such as color, texture and shape properties were extracted. This dissertation classifies the multi-instance learning algorithms under three categories. The classical algorithms of each category applied to the image retrieval. The dissertation analyses of the advantages of each type of algorithm.