Approximate Compression: Enhancing compressibility through data approximation.

摘要

The implicit noise tolerance of emerging Recognition, Mining and Synthesis (RMS) applications provides the liberty from conforming to the "correct" output. This attribute can be exploited by introducing inaccuracies to the datasets, to achieve performance benefits. Data compression provides better utilization of the available bandwidth for communication. Higher gains in compression can be achieved by understanding the characteristics of the input data stream and the application it is intended to be used for. We introduce simple approximations to the input data stream, to enhance the performance of existing lossless compression algorithms by gradually and efficiently trading off output quality. For different classes of images, we explain the interaction between the compression ratio and the output quality, time consumed for approximation, compression, and decompression. This thesis demonstrates and quantifies the improvement in compression ratios of lossless compression algorithms with approximation, compared to the state-of-the-art lossy compression algorithms.