A perceptually-based heuristic codebook design algorithm

Although the generalized Lloyd algorithm (GLA) minimizes the average squared error on the training set, the squared-error distortion measure cannot discriminate between edge and shade vectors. Moreover, the GLA is dependent on the initial codebook. These drawbacks can lead to a perceptually poorly performing codebook. A heuristic algorithm is presented which compensates for the above deficiencies to render a high-quality codebook from the perceptual point of view. This algorithm has been developed for a special vector-quantization-based image coding method, the mean/quantized residual vector quantizer (MQRVQ) method. The parameters used for tuning the algorithm depend on the distortion measure and the image vector quantization method. Thus, it lacks the generality of the GLA. In the context of MQRVQ, the algorithm has resulted in a codebook which is very close to a local optimum while providing good perceptual quality for reconstructed images. The details of this ad hoc codebook design algorithm are described. Experimental results and reconstructed images are shown.<
>