Research On Algorithms Of Video Deblocking Based On Deep Learning

With the development of information technology and application of deep learning technology,the demand for higher-quality multimedia content continues to increase.Thus technical researches on lower requirements of bandwidth and more efficient video encoding and decoding has always been hotspots in academia and industry.In recent years,deep learning technology has achieved great success in both high-level and lowlevel computer vision tasks.However,most of the tasks are based on time-independent image mediums,which means that video sequences with spatial–temporal related features are still valuable for researchers.Based on deep learning technology,this thesis mainly are divided into the following two sections: first of all,we propose a video post-processing deblocking algorithm based on temporal boosting residual networks,which innovatively utilizes the temporal characteristics of video content for multi-scale video enhancement.Besides the algorithm is able to be applied freely because of its independence to codec structure.Secondly,based on state-of-the-art technologies of network compression,we demonstrate systematic implementation of network pruning and visualization referred to Soft Filter Pruning(SFP)algorithm.And we make a certain discussion on the interpretability of neural networks.Not being constrained by the network structure of image decompression tasks,we utilize the temporal features among the sequences adequately.We propose a video deblocking framework with coarse-to-fine content enhancement and the framework integrates similar residual features among adjacent frames.Each series network makes reference to the structure of ARN(Adaptive Residual Network).The previous series network extracts coarse-grained noise features and the residual is concatenated to the next input frame,which implies a pre-filter processing.Continuously the following series network extracts more fine-grained features so as to obtain a better result.The multi-scale deblocking algorithm proposed in this thesis avoids the difficulty of training jointly with deep network structure,and achieves state-of-the-art results compared to the current similar methods.Under the same bit rate condition,it obtains 0.6dB?1.0dB improvement on average in both objective and perceptual quality.The compression of network models has always been the focus of computational theory research for neural network technology.Specifically,network interpretability and requirements of huge computing resource are important limitations to its application.This thesis implements a network compression and visualization system with considering the importance of network parameters which is based on the state-of-the-art network pruning technology.The system optimizes requirements of the storage and running time of the existing network algorithms.Briefly speaking,the convolutional features in deep networks are scored for importance(Feature Selection),then the low-importance neurons are suppressed according to the cropping ratio.Besides,to a certain content,we explore the network interpretability based on the network pruning theory and visualization technology.