Research On All-zero Block Detection For Efficient Video Coding With RDO Quantization Support

Quantization is the key module in video encoder and directly determines the distortion and bitrates.HEVC supports many intra-frame and inter-frame coding modes,which is the key to achieve higher coding performance.Rate distortion optimization is the basis of high-performance mode selection and video coding algorithm optimization.In order to achieve the rate-distortion optimization mode selection,it is necessary to calculate the encoding cost of various coding modes to achieve the minimum rate-distortion mode selection.Compared with hard-decision quantization(HDQ),rate-distortion optimized quantization(RDOQ)in HEVC brings non-negligible coding gain,however consumes considerable computations caused by exhaustive search over multiple candidates to determine optimal output level.Because the rate-distortion optimization consumes amounts of considerable computations in the process of mode selection and quantization,it is urgent to design an efficient RDOQ algorithm.Benefiting from efficient prediction in HEVC,transform blocks are frequently quantized to allzero blocks,especially for small-size blocks.Since these all-zero blocks do not carry any valid information,it is worthwhile to detect all-zero block for transform blocks to bypass subsequent computation-intensive RDOQ.Compared with RDOQ,HDQ does not use rate-distortion optimization(RDO)criteria,and directly uses simple rounding operations.The HDQ algorithm uses threshold modes to achieve all-zero block pre-decision,it will reduce the complexity of quantization algorithm.The RDOQ algorithm uses dynamic programming strategy to select the optimal quantization level from many candidate cases.Traditional thresholding based AZB detection algorithms are well-suited for HDQ quantized blocks,however miss partial optimal results in RDOQ and suffer from more or less accuracy degradation in RDOQ.Hence,we should make full use of the characteristics of RDOQ quantization to explore a more efficient all-zero block decision algorithm.So that,it achieves the best balance between computational complexity savings and quantization accuracy.This paper aims to exploit an efficient all-zero block decision for RDOQ.We explore an all-zero block decision algorithm from multiple perspectives which suits for different quantization cases.It is obtained through analysis of statistic.The main contributions are as follows.(1)Our analysis finds that HDQ and RDOQ have a higher probability of obtaining the same quantization results,which means that the RDOQ can use HDQ threshold comparison to achieve all-zero block decision.This paper firstly obtains threshold mode that suits for HDQ based on the mathematical derivation of the DCT coefficient distribution parameters.It will determine these safe transform blocks(HDQ and RDOQ quantization results are the same block);(2)We find that HDQ and RDOQ also have different quantization results.For these HDQ quantized as non-all-zero blocks but RDO quantized as all-zeros(singular samples),this paper explores a more rigorous decision algorithm.We propose an adaptive threshold mode that suits for RDOQ algorithm through knowledge analysis.The adaptive threshold mode is based on quantization parameters(QP)and it can make relatively accurate all-zero block pre-decisions for these singular samples.(3)We can't detect accurately some singular samples even with the above two methods.Thus,this paper proposes an all-zero block pre-decision algorithm based on machine learning binary classification mode.Firstly,we extract eight feature parameters that affect whether the RDOQ results are all-zero blocks.And then eight feature parameters are input into a neural network for training to obtain a binary classification mode.So that,we get high-precision RDOQ all-zero block decision through the complex neural network descripts the complex RDOQ process.The experimental results demonstrate that the proposed algorithm achieves up to 7.471% total coding computation saving with no larger than 0.064% BDBR increment on average.Moreover,the average of False Positive Rate(FPR)and False Negative Rate(FNR)are 6.5% and 6.3% respectively.