Stationary Probability Model for Microscopic Parallelism in JPEG2000

Abstract

Abstract-Parallel processing is key to augmenting the throughput of image codecs. Despite numerous efforts to parallelize wavelet-based image coding systems, most attempts fail at the parallelization of the bitplane coding engine, which is the most computationally intensive stage of the coding pipeline. The main reason for this failure is the causality with which current coding strategies are devised, which assumes that one coefficient is coded after another. This work analyzes the mechanisms employed in bitplane coding and proposes alternatives to enhance opportunities for parallelism. We describe a stationary probability model that, without sacrificing the advantages of current approaches, removes the main obstacle to the parallelization of most coding strategies. Experimental tests evaluate the coding performance achieved by the proposed method in the framework of JPEG2000 when coding different types of images. Results indicate that the stationary probability model achieves similar coding performance, with slight increments or decrements depending on the image type and the desired level of parallelism.

Bitplane Image Coding With Parallel Coefficient Processing

“…with the same wavelet filter-bank are statistically similar [35], [37], [38]. A more in-depth study on this stationary probability model can be found in [33].…”

“…The main problem of coding pass parallelism is that in order to code a coefficient in the current pass, some information of its neighbors coded in previous passes may be needed. This is addressed by delaying the beginning of the execution of each coding pass some coefficients with respect to its immediately previous pass [4], [33]. Such an elaborate strategy is not suitable for SIMD computing since each coding pass carries out different operations, which generates divergence among threads.…”

“…To do so, the probability adaptation should be carried out for all data of the codeblock, which is not possible due to the parallel processing of coefficients. Such models achieve poor performance when coding short sequences [33], so to use them independently for each stripe is not effective.…”

“…As shown in [33], this model is based on the empirical evidence that the probabilities employed to code all symbols with a context are mostly regular in the same bitplane. The probability estimates are precomputed off-line and stored in a lookup table (LUT) that is known by the encoder and the decoder, so there is no need to transmit it.…”

Strategy of Microscopic Parallelism for Bitplane Image Coding

“…The main idea is to use a fixed probability for each context and bitplane. As shown in [10], this model is based on the empirical evidence that the probabilities employed to code all symbols with a context are mostly regular in the same bitplane. The probability estimates are precomputed off-line and stored in a lookup table (LUT) that is known by the encoder and the decoder.…”