The Study On Image Annotation And Ranking In Image Retrieval

Due to the great success of text-based information retrieval, most existing image retrieval systems such as Google and Baidu are text based that judge image relevancy to the query by image tag or texts surrounding the image. But because the textual information is always noisy, the retrieval results are usually unsatisfying. There are usually two ways to improve the image search results:First, images are annotated with proper tags before retrieval, then the tags are used to retrieval images. Second, reorder of the image retrieval results so that the related images are ranked top in the list to improve search results satisfaction.Real images are often ambiguity, the traditional image tagging and classification algorithm will lead to a decline in accuracy. Multi-instance learning is a recently emerging learning framework, and has been successfully used in the task of image classification due to its outstanding representation capability of ambiguity object. The generating method of the multi-instance bag is an important factor influencing the result of multi-instance learning. This thesis focuses on the analysis of multi-instance image classification and proposes an entirely new image multi-instance bag generation method. At the same time, to improve the generalization ability of the classifiers, we propose a novel selective ensemble learning method. Our method applies a discrete space optimization, and thus is very fast. On the other hand, the image itself is a multi-modal object. Most image ranking methods don’t fully explore the interaction between multiple image modalities such as textual and visual information. This thesis comprehensively considers the mutual relationship between the multiple image modalities and proposes a novel image ranking framework.Three aspects of work are included in this paper to improve the search performance:1. Analyze traditional bag generation methods in multi-instance image classification and propose a new bag generation and image classification method. In the proposed approach each image is modeled as a Gaussian mixed model and each component Gaussian distribution is one instance in the multi-instance bag, which can obviously preserve more image information than traditional vector instance.2. Analyze the existing ensemble learning methods and propose a novel multiset based selective ensemble method, which is a discrete space optimization, and is very fast.3. Analyze the existing image ranking method and propose a novel image ranking framework which will fully explore image multi-modal relationships. In the proposed framework, the retrieved result set is modeled as a multi-graph, where each image is a node with multimodal attributes (textual and visual features) and the parallel edges between nodes measure both image intra-modal and inter-modal similarities. Random walk is employed to compute the final ranking result.Finally, experiments are conducted to demonstrate the effectiveness of the presented models on Corel image dataset, UCI machine learning repository and Web Queries dataset respectively.