An efficient predictive coding of integers with real-domain predictions using distributed source coding techniques

By exploiting the commonly observed Laplacian probability distribution of audio, image, and video prediction residuals, many researchers proposed low complexity prefix codes to compress integer residual data. All these techniques treated predictions as integers despite being drawn from the real domain in lossless compression. Among these, Golomb coding is widely used for being optimal with non-negative integers that follow geometric distribution, a two-sided extension of which is the discrete analogue of Laplacian distribution. This paper for the first time presents a novel predictive codec which treats real-domain predictions without rounding to the nearest integers and thus avoids any coding loss due to rounding. The proposed codec innovatively uses the concept of distributed source coding by replacing the reminder part of Golomb code with the index of the coset containing the actual value.