A hybrid image data compression scheme using quaternary decomposition and vector quantization of DCT coefficients

This research investigates a hybrid data compression scheme applicable to both single frame color images and color image sequences. It uses a form of classified vector quantization (VQ) based upon the quaternary decomposition of image blocks into 32 x 32(x4), 16 x 16(x4), 8 x 8(x4) or 4 x 4(x4) sub-blocks before either a 2-D or 3-D discrete cosine transform (DCT) is used to encode the blocks. The coefficients of the DCT's are then zonal sampled and vector quantized. The codebooks for the VQ are generated by either a random or a clustering process for the single frame images, while only the clustering process is used for the image sequences. Each of the initial codebooks is optimized by both the Linde-Buzo-Gray (LBG) method and a simulated annealing (SA) method. In addition a method of selective multistage vector quantization (SMVQ) is investigated in which only those vectors not matched well enough by a codebook vector are allowed a second stage of VQ.

Results indicate that the SA algorithm does indeed produce better results than the LBG method, although the effects are masked by the combination of vectors for the final reconstruction. The compression scheme produces very good results for single frame images at a bit rate of.858 bpp, while the image sequences are processed to a similar quality at a bit rate of.395 bpp. When the SMVQ method is used the bit rates increase by approximately.15 bpp in both the single frame and sequence cases, with improved quality in the reconstructed images, especially in the lower quality cases.