Complexity Reduction and Control Techniques for Power-Constrained Video Coding

The ever-increasing demand for multimedia content imposes a big challenge for video compression. Today, smartphones are widely popular and in use. The high-quality cameras and high-resolution screens on these devices have elevated the user expectations for a higher resolution and higher quality video content. As a result, more High Definition (HD) and Ultra High Definition (UHD) content are being shared by users. The High Efficiency Video Coding (HEVC/H.265) was introduced as the latest video coding standard by ITU and MPEG, to cope with the high bitrate of this large content. It can provide almost twice the coding efficiency, compared to its predecessor, H.264/AVC. However, the encoding process for this standard includes algorithms with a higher computational complexity, due to the newly introduced coding tools, and a higher precision for the existing tools. As a consequence, HEVC encoding is power demanding, and thus hard to employ especially on battery-powered devices. Furthermore, many applications e.g. video streaming, cloud gaming, and power-constraint video coding, are sensitive to the coding delay (or power), and their tolerable delay (or processing power) varies through time. For this reason, not only reducing the complexity of video encoding is important, but a power-adaptive mechanism is crucial to enable the encoding within the defined power/delay quota.

To address these issues, this thesis investigates the complexity of HEVC encoding and presents novel complexity reduction and complexity control algorithms for HEVC encoding. These contributions can be categorized in three main parts. The first contribution of this thesis considers the high complexity and power consumption of motion estimation in HEVC. Specifically, the baseline encoding algorithm is analyzed from a memory access point of view, which contributes heavily to the total power consumption of video encoding. A simple yet effective approach is presented that replaces the first step of the search for finding the best starting point, and also adaptively reduces the search range. This method reduces the memory access and encoding time, with negligible loss of coding efficiency.

The second contribution of the thesis is fast intra prediction. As the complexity of intra coding has particularly increased in HEVC, more investigations have been dedicated to this module. Two novel methods have been proposed that accelerate the intra prediction through fast texture analysis. The first approach adopts filters of a Dual-Tree ComplexWavelet Transform(DT-CWT) to estimate the texture direction of each intra block. The second approach exploits the potentials of internal tools integrated in the HEVC engine, i.e. the planar filter and the entropy engine, to prune unnecessary computations.

The third contribution of this work is a machine-learning driven approach for controlling the complexity of HEVC encoding, through fast Coding Unit (CU) partitioning. To this end, a feature set is designed for CU partitioning, which uses DTCWT for advanced texture analysis. Then an adaptive classification approach is presented that decides the termination or skipping of eachCUdepth. The performancecomplexity of this classification scheme can be adjusted through a set of thresholds. Finally, the complexity control problem is modeled and solved as a constraint optimization problem, where the loss of the coding efficiency is minimized while the coding complexity is set to meet the target level.

The effectiveness of the proposed algorithms is verified through extensive experiments over the video dataset suggested in the common test conditions of the standardization committee.