The Digital Signal Processing Lab @ UCSD


Thesis Title

Gaussian Mixture Models in Compression and Communication

Thesis Abstract

Gaussian mixture models (GMMs) are employed in developing low complexity fixed-rate and variable-rate compression schemes and in the design of joint source-channel decoding architectures.

In fixed-rate compression, the problem of source coder design is formulated as one of estimating the probability density function (PDF) of the source using a GMM and then designing a low complexity fixed-rate compression scheme based on the model. An efficient quantization scheme using transform coding and bit allocation techniques which allows for easy and computationally efficient mapping from observation to quantized value is developed. An attractive feature of this method is that source encoding using the resultant code-book involves very few searches and its computational complexity is minimal and independent of the rate of the system. The flexibility of the proposed compression framework allows for important extensions using lattice vector quantization and in recursive coding.

In variable-rate compression, practical algorithms for lossy coding of arbitrary sources using a combination of lattice quantization, GMMs of source PDFs, and arithmetic coding is proposed. At high rates and with an accurate GMM for the source, the performance degradation is shown theoretically and experimentally to be only that of the space-filling loss of the lattice chosen.In joint source-channel decoding, GMMs are used along with high resolution theory in the design of decoding architectures for erasure and AWGN channels.

In erasure channels, the GMM of the joint PDF of successive source frames is used to process the channel decoder output over erasure channels. The performance of two decoding strategies, namely, Maximum Likelihood decoding (ML) and Minimum Mean Squared Error decoding (MMSE) is shown to provide significantly better performance than prediction based schemes.

In AWGN channels, an iterative decoding procedure is proposed that works with any channel code whose decoder can provide extrinsic information on each source encoded bit. The source decoder uses the GMM of the joint PDF and the channel decoder output to provide a priori information back to the channel decoder. Experimental results showing improved performance are provided in the application of speech spectrum parameter compression and communication.

Year of Graduation: 2003