Approximated transform and quantisation for complexity-reduced high efficiency video coding

Abstract

The transform-quantisation stage is one of the most complex operations in the state-of-the-art High Efficiency Video Coding (HEVC) standard, accounting for 11–41% share of the encoding complexity. This study aims to reduce its complexity, making it suitable for dedicated hardware accelerated architectures. Adopted methods include multiplier-free approach, Multiple-Constant Multiplication architectural designs, and exploiting useful properties of the well-known Discrete Cosine Transform. Besides, an approximation scheme was introduced to represent the original HEVC transform and quantisation matrix elements with more hardware-friendly integers. Out of several derived approximation alternatives, an approximated transform matrix (T16) and its downscaled version (ST16) were further evaluated. An approximated quantisation multipliers matrix (Q) and its combination with one transform matrix (ST16 + Q) were also assessed in HEVC reference software, HM-13.0, using test video sequences of High Definition (HD) quality or higher. Their hardware architectures were designed in IEEE-VHDL language targeting a Xilinx Virtex-6 Field Programmable Gate Array technology to estimate resource savings over original HEVC transform and quantisation. T16, ST16, Q, and ST16 + Q approximated transform or/and quantisation matrices provided average Bjøntegaard-Delta bitrate differences of 1.7%, 1.7%, 0.0%, and 1.7%, respectively, in entertainment scenario and 0.7%, 0.7%, -0.1%, and 0.7%, respectively, in interactive scenario against HEVC. Conversely, around 16.9%, 20.8%, 21.2%, and 25.9% hardware savings, respectively, were attained in the number of Virtex-6 slices compared with HEVC transform or/and quantisation. The developed architecture designs achieved a 200 MHz operating frequency, enabling them to support the encoding of Quad Full HD (3840 × 2160) videos at 60 frames per second. Comparing T16 and ST16 with similar designs in the literature yields better hardware efficiency measures (0.0687 and 0.0721, respectively, in mega sample/second/slice). The presented approximated transform and quantisation matrices may be applicable in a complexity-reduced HEVC encoding on hardware platforms with non-detrimental coding performance degradations.