A Memory Source Model Researched On Video Coding Quantization Algorithm

With development of video technology,a new generation of encoding standard H.264/H.265 emerged as the key role of quantization directly determines the performance of video.Most quantization uses memoryless source assumption HDQ,which is easy to realize with low complexity.Research shows that the assumption of memoryless source is not true and the performance and model accuracy need to be improved.Consideringcorrelation of coefficients,SDQ improves video quality greatly,but the high computational complexity is not suitable for parallel processing.Based on this,we improves soft and hard decision quantization algorithm,and takes into account the two advantages of the two kinds of algorithms,so as to ensure that the parallel processing does not increase the computational complexity and introduces the coefficient correlation.In order to improve quantization performance,an improved adaptive offset HDQ algorithm with memory source model is studied.Aiming at the loss of quantization rate-distortion performance caused by fixed HDQ offset,an improved HDQ algorithm based on DCT coefficient adaptive offset adjustment is studied in this paper.Using off-line modeling,the function model of offset and quantization parameter DCT coefficient distribution parameters is constructed.Compared with the original fixed offset model,the improved algorithm adaptively adjusts the quantized offset and effectively improves the rate-distortion performance.Aiming at the independence of HDQ coefficients and combining with theory of SDQ context rate entropy coding,a new rate adaptive HDQ improved algorithm is proposed by adding rate factor to the previous improved HDQ model.We estimate the rate self-information that may be consumed when quantization results are different online,and analyze the fine-tuning effect of the rate self-information on the quantization offset of the current coefficient.Based on the maximum positive decision,we analyze the effect of the rate self-information on the quantization offsetof the current coefficient.A function model of four quantization parameters of offset and bit-rate self-information is established by off-line minimum error probability.The experimental results show that the improved HDQ with bit rate can improve the performance of HDQ rate distortion compared with the HDQ with independent coefficients.Compared with SDQ,it has lower computational complexity and is more consistent with the actual situation of the correlation between memory sources.In view of the large number of parameters affecting quantization,complex process of traditional mathematical modeling and the slow development of the research model,neural network is applied to the modeling of video quantization algorithm.According to relationship between quantization parameters and quantization offset,four important parameters affecting quantization are selected,and a HDQ offset selection algorithm based on neural network is designed.The best number of hidden layers,nodes and the iterations are designed.Training results converge to the range of preset precision,which provides a train of thought for the study of quantization algorithm parameter selection.The experimental results show thatimproved hard decision algorithm based on neural network is convenient and simple.