Abstract

In this paper, the incremental codebook generation process, which is a technique for representing a database of images as a single codebook, that captures the content of all the images is proposed. Vector quantization (VQ) is used for creating the codebook of the image. The main problem with VQ is the size of the training sequence that is used to generate the global codebook. This paper explains a method, where the local codebooks are generated for each image. Then, the locally global codebook for the images is computed from the local codebooks by the incremental process. In order to handle large number of images of similar classes, a new method of incrementally creating locally global codebook based on hierarchy of image classes is also explained. It gives an automated representation of the image database compared to the previous work of selecting a random set of representative images, to compute the global codebook. The locally global codebook is used to cluster the images using the encoding distortion. This codebook helps to match images, and retrieve images that are similar to it by comparing the encoding distortion. The precision and recall curve and mean average precision gives better results than the previous works. The classification error is reduced as the distortion is always the smallest for similar images.