Content area
Optimal scalar and vector quantizers that can be used on the trans- formed image sequence frames are examined and a new method for vector quantization is proposed. Quantizers are essential components of any compression system and they are the main contributors to the reconstruction error incurred at the decompression. As image transformation, Discrete Cosine Transform (DCT) is chosen, since it is a good approximation to Karhunen-Loève transform which perfectly decorrelates its input data.
For scalar quantization in DCT domain, nonuniform Max quantizers tailored for the distribution of DCT coefficients of transformed image data are used and integer bit allocation method of marginal analysis is chosen. Further, incorporation of human visual system properties to this optimal scalar quantization scheme is extensively discussed and simulations are performed.
For vector quantization in DCT domain, classified vector quantization (CVQ) techniques are examined and a novel scheme named as Classifier Constrained Vector Quantization (CCVQ) is introduced. This technique uses some elements of the input vector as class features and jointly optimizes the VQ codebook and the classifier, using Lagrangian methods. Therefore, it generates sub codebook for each class with optimal sizes and minimum distortion. This new method is tested experimentally and shown to have a very short execution time, although its PSNR performance is very close to full-search VQ techniques.
In order to choose the best quantization policy, optimal scalar quantization and vector quantization are compared both theoretically and practically.
As a result, vector quantization is confirmed to be superior over any scalar quantization scheme.
Finally, the intraframe coded image sequence frames with the above techniques are interframe coded to reveal the potential of compression in temporal direction.