Study On Bag Of Visual Word Based On Local Invariance And Color Constancy

Abstract:With the explosive growth of images and the arrival of the age of big date, there is a higher requirement for image classification or retrieval technology nowadays. The research of key technologies for image classification or retrieval, especially the research of image representation method is significant for the efficient use of images under the age of big date. The conventional image representation model based on global features, such as color or texture features, can hardly represent the content and meaning of images, especially when the region of interest is local features. Therefore, the research of Bag of Visual Word based on local features is quiet valuable for image classification or retrieval technology.The key technologies of the study of Bag of Visual Word are the method of local features and the mapping method from feature to visual word, and this thesis conducts deep researches and experiments on these two key technologies.On the basis of the analysis of the detection operator and description operator of local features, the research is mainly focused on the SIFT algorithm and the ORB algorithm. The experimental results of local features show that the FAST of ORB has an ideal detectability and a higher speed; the binary code string descriptor of ORB is much more superior to SIFT descriptor in the aspects of storage and calculating, while they have similar performance. The local feature operator based on gray scale has a poor illumination robustness.Considering that the local features are sensitive to illumination, the color constancy theory is studied. On the basis of the combination of ORB algorithm and color constancy, a kind of binary code string descriptor with illumination robustness:the CCORB algorithm is proposed. Experiments show that CCORB is much more advantageous over ORB in illumination robustness.According to the binary code string descriptor that CCORB adopted, a new mapping method of visual words with multiple weights and the corresponding visual dictionary training method are presented. Furthermore, the histogram representation model of visual words is provided. Experimental analysis shows that:the Bag of Visual Word presented in this thesis can provide a better discrimination in image classification.The research achievements of this thesis provide a theoretical basis for the application of Bag of Visual Word in image classification or retrieval.