Fast image search on a VQ compressed image database

Authors: MEHMET YAKUT

Abstract: A fast and efficient image search method is developed for a compressed image database using vector quantization (VQ). An image search on an image database requires an exhaustive sequential scan of all the images, given the similarity measure. If compressed images are dealt with, images are decompressed as an initial operation and then the previously mentioned exhaustive search is performed using the predetermined similarity measure. If the images in the database are compressed using VQ, the image search process is reduced to codebook index match tests. A pixel by pixel similarity test of two images computationally costs too much. This bottleneck is overcome by using VQ, where the similarity test of the two image block is performed by a precalculated distortion lookup table. The same is valid for the object search in the image database. The object image is vector quantized first; then the index map of the object image is scanned over the entire index area of the compressed image database. Significant image search speed gains on the VQ image database are obtained. Results show that the VQ compressed image search is faster than a sequential search, and compressed and decompressed JPEG search. Actual speed gain obtained here depends on the application area and required image quality for the database.

Keywords: Image search, image compression, vector quantization, compressed recognition

Full Text: PDF