Locally-adaptive perceptual quantization without side information for DCT coefficients

Existing JPEG/MPEG compliant perceptual coding methods do not fully exploit the local variation of perceptual masking thresholds. This work demonstrates that, for natural images with the same perceptual quality, the first-order entropy of the quantizer outputs can be reduced by 15 to 40 percent when optimal locally-adaptive perceptual quantization is used. Locally-adaptive perceptual quantization requires the perceptual thresholds to be available at the decoder. However, transmitting the thresholds is prohibited by the large amount of side information required. In this paper, a DCT-based method is introduced that performs locally-adaptive perceptual quantization without side information based on estimates of the perceptual thresholds from the already quantized data. Since no additional side information is required, this method is compliant with the bit stream syntax of the above mentioned standards.