Video Coding Quantization Prediction And Multi-module Relevancy Measurement

Video coding is the process of mapping digital video signal to a smallerrange, and at the same time ensures that videos can be restored as normal visualeffects after decode. Choosing an optimal quantization parameter (QP) is one keytechnique in rate control, which is able to reduce visual redundancy and code lengthwhile balancing the image quality and encoding complexity. H.264/AVC’s ratecontrol algorithm is the most widely used among varieties of code rate controlschemes, it models the video quantization processes and changes the quantitativevalue according to the actual image dynamically to trade off between code length andimage accuracy for achieving best integral effect.But the rate control scheme is imprecise in controlling sequences with highmotion and scene cut, and too sensitive to bandwidth fluctuation. In this paper, we doresearch on the quantitative parameter prediction model in rate control, and improvethe original model combining with neural network. The method is to establish athree-layer BP network using the information of encoded frame to adjust the QP ofcurrent frame. After training, the model predicts QP to quantize the coded framedirectly. The experiment shows that，after the application of BP neural network’slearning ability and robustness, this scheme reduces the PSNR fluctuation and getmore smooth video quality in the guarantee of rate accuracy at the same time.After improving the traditional rate control algorithm, we aim at couplingphenomenon in multiple-module algorithm design. The sum of the R-D performancevariation due to two individual module modifications is not linear with the R-Dperformance variation due to simultaneous modifications of these two modules. Sothe next work is how to measure the inter-influence among modules.Due to the lack of systematic investigation on mulit-module algorithm design,this paper attempts to measure the inter-relevancy between modules. According tothe algorithm characteristics and traditional standard definition, we group thecustomizable algorithms accounting for the application targets into four typicalmodules: video pre-processing, rate control, motion estimation and mode decision, and gives quantitative measure method for inter-relevancy level and importancefactor. X264is used as the verification platforms, and experimental results are givento verify the proposed idea and model. This work is instructive and meaningful toguide multi-module joint algorithm optimization.In order to improve the video rate distortion performance as far as possible, alsoweigh the coding complexity, it is needed to coordinate these undetermined modules.Therefore according to relevance metrics, this paper uses the two importance factorswe get to improve the video coding optimization method. With the designed schemeto select optimal encoding parameters, we can improve coding performance asgreatly as possible as the global optimum, at the same time its complexity is less thanhalf of the latter.