Research On Bit Rate Prediction Based On Video Content Feature For 4K Sequences

As video has become one of the most popular ways to get information,people tend to show higher demand on viewing experience.4K Ultra High Quality video ensures a better user experience for its resolution of 3840×2160 and delicate quality.However,4K video requires higher bit rate and larger storage space compared with common videos.Therefore,how to balance the need of good video quality with saving bit rate is a key to 4K industry.Under this background,this thesis focuses on bit rate prediction for 4K sequences based on video content feature.Both subjective and objective video quality assessment are considered during our research.Nowadays relative fields are still lack of available research-free database of 4K video sequences with subjective assessment results.Hence,to fill the gap,subjective video quality tests(including 4K,HD and 720 p sequences)are carefully conducted.After analysis on relationship between video quality and bit rate,we notice that video content has great impact on bit rate given a fixed video quality,which provides a basis for choosing video content to predict bit rate.Usually,videos need to be encoded repeatedly to predict bit rate,while this thesis focuses on no-reference objective video quality assessment model in a non-encoding way to reduce the computation complexity.We first make a comparison on several no-reference models on our 4K and HD sequences,and then we calculate several video content and distortion features.As a result of Principal Component Analysis,content features are chosen as video features.At last we modify the model by considering the content features extracted before.Results show that our model has a better performance with a fact that the Pearson Correlation Coefficient of our model increases by 21.76%.Optimization of the whole bit rate prediction approach is also conducted.We use clustering analysis to tighten the sequences with similar content features,and then we predict bit rate on data of each cluster instead of the whole database.Results show that the prediction performance increases by 18.97%.Taking sequence ‘Marathon' as an example,the encoding bit rate is reduced by 69.40% compared with 15000 kbps which is the common setting for encoding 4K sequences in the industrial world.Besides,we notice that the inflection point of “quality-bit rate” curve may increase the capacity of saving bit rate.Therefore,we propose a rule of selecting which bit rate as the optimized one.This thesis accomplishes the goal of saving bit rate when good video quality is required.