Research On Data-driven Algorithms For High Efficiency Video Coding

As the development of multimedia,the study on video coding approaches has becoming more and more popular.Since the newest video coding standard named High Efficiency Video Coding(HEVC)has a promising compression performance and a great compression compatibility,it has become the most popular video coding standard.Similar with previous standards,HEVC also adopts the hybrid video coding framework with traditional filters to accomplish prediction and enhancement.Despite of low complexity,these traditional filters cannot handle various video contents well,which limits the coding performance.To further improve the coding performance,we utilize the characteristics of video coding to propose a data-driven based high efficiency video coding algorithm.The major contents of the dissertation are summarized as follows:1.A multi-scale feature fusion framework for HEVC loop-filter algorithm.Since HEVC adopts the quadtree partition structure,the introduced compression artifacts have multi-scale feature similarity.Besides,since video coding is sensitive to realtime capability,the loop-filter algorithm should have a low model and time complexity.Based on these two characteristics,we proposed a new HEVC loopfilter algorithm.First,we utilize the recurrent residual structure to capture the multiscale similarity of compression artifacts in HEVC.Then,we share the filter parameters among different convolutional layers to control the model complexity and ensure the enhancement performance.2.A multi-task learning framework for HEVC interpolation.To improve the precision of motion vector and reduce the prediction error,HEVC adopts multi-tap interpolation filters to prediction the fractional pixels.Since the frame to be interpolated is reconstructed through coding and decoding,it is degenerated with irreversible compression artifacts.Therefore,the interpolation algorithms in HEVC should additionally consider the noise characteristics of the frame to be interpolated and void propagating the noise into the fractional pixels.Based on these,we proposed a multi-task learning framework for HEVC interpolation algorithm.Our approach first utilizes the multi-task learning framework to capture the noise characteristics in the frame to be interpolated,and provide the prior for subsequent interpolation.Then,a deconvolutional layer is utilized to generate fractional pixels unitedly to reduce the algorithm complexity.The experimental result shows that,compared with the original fractional interpolation filter in HEVC,our algorithm achieves 5.1%,4.2% and 1.8% BD-rate reduction under Low-Delay P,Low-delay B and Random Access configuration.3.An asymmetric fusion framework for blind video quality enhancement.Since compressed videos adopt similar coding structures and prediction modes under different quantization parameters,the compression artifacts under different quantization parameters also have feature similarity.Besides,the study shows that under-enhancement(utilizing the network trained with low quantization parameter to enhance the video compressed with high quantization parameter)has a higher enhancement performance than over-enhancement(utilizing the network trained with high quantization parameter to enhance the video compressed with low quantization parameter).Based on these two enhancement characteristics,we propose a blind video quality enhancement approach.First,a progressive feature extraction structure is proposed to extract the enhancement features of different quantization parameters unitedly.Then,based on the different performance of overenhancement and under-enhancement,the enhancement features are asymmetrically fused to achieve an optimal enhancement performance.The experimental result show that our approach achieves over 0.5 d B PSNR increase under different coding configurations and different quantization parameters.