Assorted Color Spaces to improve the Image Retrieval using VQ Codebooks Generated using LBG and KEVR

The paper presents performance comparison of image retrieval methods based on texture feature extraction using Vector Quantization (VQ) codebook generation techniques like LBG and KEVR (Kekre‟s Error Vector Rotation) with assorted color spaces. The image is divided into non overlapping blocks of size 2x2 pixels (each pixel with red, green and blue component). Each block corresponds to one training vector of dimensions 12. The collection of training vectors is called a training set. The texture feature vector of the images are obtained from the most popular VQ algorithms LBG and KEVR applied on the image training set and codebooks of size 8, 16, 32, 64 128, 256 and 512 are generated. These codebooks are the feature vector set for Content Based Image Retrieval (CBIR). The results are obtained using six different color spaces such as RGB, LUV, YCgCb, YCbCr, YUV and YIQ. For experimentation, the generic image database having 1000 images is used. From the results it is observed that KEVR based CBIR shows better performance over LBG based CBIR. Overall in all codebook sizes KEVR in YUV color space gives the best results with higher precision-recall crossover point values; closely followed by YCbCr color space.