ABSTRACT: Context-adaptive binary arithmetic coding is a widespread technique in the field of image and video coding. Most state-of-the-art arithmetic coders produce a (long) codeword of a priori unknown length. Its generation requires a renormalization procedure to permit progressive processing. This paper introduces two arithmetic coders that produce multiple codewords of fixed length. Contrarily to the traditional approach, the generation of fixed-length codewords does not require renormalization since the whole interval arithmetic is stored in the coder's internal registers. The proposed coders employ a new context-adaptive mechanism based on variable-size sliding window that estimates with high precision the probability of the symbols coded. Their integration in coding systems is straightforward as demonstrated within the framework of JPEG2000. Experimental tests indicate that the proposed coders are computationally simpler than the MQ coder of JPEG2000 and the M coder of HEVC while achieving superior coding efficiency.

The arithmetic coders introduced in this work are implemented in Java and incorporate procedures to allow their use in a large variety of applications (see the manuscript for a detailed explanation). The implementations employed to generate the experimental results of this work are left freely available: