A model-adaptive universal data compression architecture with applications to image compression

Abstract

In this thesis, I designed and implemented a model-adaptive data compression system for the compression of image data. The system is a realization and extension of the Model-Quantizer-Code-Separation Architecture for universal data compression which uses Low-Density-Parity-Check Codes for encoding and probabilistic graphical models and message-passing algorithms for decoding. We implement a lossless bi-level image data compressor as well as a lossy greyscale image compressor and explain how these compressors can rapidly adapt to changes in source models. We then show using these implementations that Restricted Boltzmann Machines are an effective source model for compressing image data compared to other compression methods by comparing compression performance using these source models on various image datasets.