Two-Stage Golomb - Context-Adaptive Binary Arithmetic Coders Coding in Lossless Image Compression

In this paper, a broad view of the individual stages of data modeling and encoding in the field of lossless image compression was presented. The main emphasis was placed on encoding prediction errors with the assumption of its geometric distribution. Basic and less common techniques of predictive data modeling for improving prediction efficiency by better fitting to a one-sided probability distribution were discussed. Among them, authors' own mechanism Conditional Move To Front (CMTF), which can be useful for encoding images with high variation of input data, was described. Additionally, an original two-stage mechanism for efficient prediction error encoding (used in three codecs of different computational complexities: Multi-ctx 2, 7-ctx MMAE, and Blend-28), which uses adaptive Golomb code at the initial stage and passes its binary output to context-adaptive binary arithmetic coders (CABAC), was described in detail. Separate coders for prediction error bit signs and prediction coefficients (often forming large block in header data) were also introduced. In particular, the authors focused on the important role of correct contextual division, when sections with different characteristics of their nearest neighborhood are grouped into separate classes to compress the data within each of them as efficiently as possible. From experimental studies, it was concluded that the minimum mean absolute error (MMAE) method is superior to the minimum mean square error (MMSE) method in determining linear prediction models, especially for images with low noise level. Connecting the mechanisms known from the literature with authors' ideas in the Blend-28 codec enabled to increase compression efficiency in comparison to the modern and popular WebP codec by achieving an approximately 11% shorter bit average.