Image vector quantization using a two-stage self-organizing feature map

This paper presents a new approach to classified vector quantization in the discrete cosine transform domain (DCT/CVQ) for image compression. While most existing DCT/CVQ methods determine class features through experiences or by studying the properties of the DCT, the proposed method attempts to extract actual class features from training images utilizing a neural network model, referred to as the self-organizing feature map (SOFM). The codebook of each class is also designed using SOFM after allocating coding bits to each class with the BFOS algorithm. In the experiments using monochromatic benchmark images, the proposed approach provided 1.07 dB similar to 1.57 dB higher peak signal-to-noise ratios, (PSNRs) than the JPEG baseline system for training images at comparable bit rates. For other benchmark images, the approach improved the PSNR by up to 0.41 dB compared with the JPEG baseline system, and by up to 0.38 dB compared with an existing DCT/CVQ method that uses the Linde-Buzo-Gray (LEG) algorithm for codebook design, depending on the bit rate used.