Automatic Image categorization is a universal problem in area of Content-based image retrieval (CBIR). The goal of automatic image categorization is to find a mapping between images and the predefined image categories. The difficulty of this problem is that how to describe image content and incorporate low-level features into semantic categories. As a solution, we propose a Principal component analysis (PCA) based approach. This approach assumes that the images in the same semantic category have the similar spatial distribution of color components and treats the images in the same category as a linear combination of a fixed set of dominant color blocks with special textural information. A three-step algorithm is designed: (1) extracting Dominant colors (DC) of images, which describe the major color information in an image; (2) Establishing a feature space based on DC blocks and its textural information; (3) using PCA to reduce dimensionality of feature space and using the basis vectors to categorize images. An experimental database containing nine categories including cars, flowers, houses, portraits, fish, bark, sunshine, leaves and fresco is constructed to test the algorithm based on our image categorization approach. The results show that this approach is effective and a reasonable compromise between accuracy and speed in practice.
展开▼
