Feature Coding And Its Applications To Image Categorization

Image Categorization is a basic problem in the field of computer vision. Generally, a typical system of image categorization utilizes the bag-of-features model, which consists of three steps: feature extraction, feature coding, classifier design. Feature coding, which serves as a connecting link in a system of image categorization, imposes great influence on recognition performance. In this thesis, some methods of feature coding are researched in order to improve the recognition accuracy. The main work and innovations of the thesis include two aspects:1. A classified sampling strategy based sparse coding spatial pyramid matching is proposed.Generally, the features of each image category vary in number and are uneven in the feature space. Compared with the random sampling, the proposed method reduces the above influence effectively. As a result, a more descriptive codebook is constructed and the descriptive ability for the images is improved. Many experimental results demonstrate that the recognition accuracy of the proposed method is obviously enhanced. Specially,the recognition rate is remarkably improved when the number of image categories is great and the training set is small.2. A coding combination of multiple features suitable for scene recognition is proposed,and the selection of the corresponding classifier is researched.Aiming at scene recognition, the proposed method takes the complexity of scene recognition into consideration, so it describes both the appearance and structure of the scenes by the coding combination of multiple features. Then the classifier suitable for scene recognition is selected, thus the recognition ability of scene images is further improved. Experimental results show that the proposed coding combination of multiple features and the selection of the suitable classifier can improve the scene recognition accuracy.