HEVC Fast Intra Coding Based On DCT Approximation And Early CU Mode Decision

Video is a more and more popular information format.As the development of video acquisition,transmission and display technology,video tends to be more complex and with higher resolution.Video coding technology aims to pass more pixels information under limited channel width.In order to encode more complex video,the computations per pixel and the pixels which are need to be dealing with both increase exponentially.HEVC as the successor of H.264,after coding according to it,the coded video which has the same quality as original only need half of the bit-rate.The high efficiency is supported by high computations.HEVC increases 5-10x computations compared to H.264.It cause more coding time in the same condition.As the developing of mobile computing technology,It's urging to design low power HEVC encoder while maintaining high efficiency.Quad-tree based recursive partition mode decision,RDO process,different combination of CU/PU/TU mode and other new features make up the majority of HEVC computation.Transform coding based on DCT dominate the distortion calculation of RDO.This paper proposes a set of method to mainly optimize above processes in order to reduce the computations of HEVC intra coding and accomplish a fast,low power,low hardware cost encoder.Mainly contributes on the following points:1,proposes a butterfly structure to calculate integer DCT.2,utilizes the energy focusing feature of DCT and drop the high frequency coefficients.Design the hardware reusing memory for transpose file.3,uses edge energy and corner energy to do early CU mode decision.4,uses CNN to solve situation which 3 can't deal with and proposes a high efficiently training scheme of CNN.After implement in HM15 platform,our DCT approximate algorithm gets a 1.03%BD-Bitrate augment while saving 15.9%coding time.Further realizing the DCT VLSI module based on TSMC 90nm,our module gets a 12.7 power consuming and 311MHz max clock speed in the worst condition.We save 71.9%hardware cost and 70.2%power consuming compared to the original design.Our CNN based fast CU partition mode decision algorithm saves 59.7%coding time with a 2.40%BD-Bitrate increase.It will be more powerful with the development of CNN hardware accelerator.