A study of bit allocation for Gaussian mixture model quantizers and image coders /
Denis Tran · eScholarship@McGill (McGill) · 2005
This thesis describes different bit allocation schemes and their performances when applied on coding line spectral frequencies (LSF) using the GMM-based coder designed by Subramaniam and a simple image transform coder. The new algorithms are compared to the original bit allocation formula; the Pruning algorithm used by Subramaniam, Segall's method and the Greedy bit allocation algorithm using the Log Spectral Distortion and the Mean-Square Error for the LSF quantizer and the Peak Signal-to-Noise Ratio for the image coder. First, a Greedy level allocation algorithm is developed based on the philosophy of the Greedy algorithin but, it does so level by level, considering the best benefit and bit cost yielded by an allocation. The Greedy level allocation algorithm is computationally intensive in general, thus we discuss combining it with other algorithms to obtain lower costs. Second, another algorithm solving problems of negative bit allocations and integer level is proposed. The level allocations are to keep a certain ratio with respect to each other throughout the algorithm in order to remain closest to the condition for lowest distortion. Moreover, the original formula assumes a 6dB gain for each added bit, which is not generally true. The algorithm presents a new parameter k, which controls the benefit of adding one bit, usually set at 0.5 in the high-rate optimal bit allocation formula for MSE calling the new algorithm, the Two-Stage Iterative Bit Allocation (TSIBA) algorithm. Simulations show that modifying the bit allocation formula effectively brings about some gains over the previous methods. The formula containing the new parameter is generalized into a, formula introducing a new parameter which weights not only the variances but also the dimensions, training the new parameter on their distribution function. The TSIBA was an a-posteriori decision algorithm, where the decision on which value of k to select for lowest distortion was decided after computing all distortions. The Generalized TSIBA (GTSIBA), on the other hand, uses a training procedure to estimate which weighting factor to set for each dimension at a certain bit rate. Simulation results show yet another improvement when using the Generalized TSIBA over all previous methods.