Object recognition and near-duplicated image retrieval have been the classical and challenging problems in computer vision. The extensive applications in human-computer interface, intelligent robot, digital media retrieval, etc. make the relevant researches very meaningful and helpful. In recent years, the bag of visual words model has attracted much attentions due to its simplicity and robustness to the environmental noise. This thesis mainly emphasizes on the discussion about the bag of visual words model, and focuses on its applications and relevant researches upon object recognition and near-duplicated image retrieval. Besides, we will analysis some existing problems in these two fields under the bag of visual words model, and accordingly provide our novel solutions. The main contents and contributions of this thesis include: 1. A comprehensive review for the image representation methods under the bag of visual model is presented, as well as the discussion about the respective advantages and disadvantages of each proposed methods. And the emphases are also placed upon each specific operation flow in object recognition and near-duplicated image retrieval under the bag of visual words model. 2. According to the problem in object recognition that the image representation will suffer from the ignorance of the spatial relationship between the local key-points, we propose the expanded bag of visual words presentation for object recognition. In this method, the classical bag of visual words representation is updated based on the Query Expansion algorithm using the explored mutual co-occurrence relationship between the visual words, and we demonstrate the significant improvement in the robustness and discriminativeness of the novel representation. 3. The problem in near-duplicated image retrieval based on the bag of visual words is uncovered that the matching key-points pairs could not always successfully be assigned to the same visual word. Accordingly, another novel algorithm is proposed in the thesis to improve the matching probability of the relevant images, and thus refine the final retrieval results. 4. We implement a real-time CD retrieval system based on the hierarchical bag of visual words model, the system is completed upon the flatform of Windows-PC and developed by Visual C++ 2008. The real-time(less than 1 sec) and accurate retrieval results can illustrate the success of the system. In a word, in this thesis, we have made a lot of fruitful...
修改评论