Research On Image Semantic Coding Algorithm Based On Rate-Distortion Theory

With the increasing popularity of 5G networks and the rapid development of 6G networks,high spectral efficiency and highly reliable information transmission poses challenges to traditional communication technologies under the guidance of classical information theory.At the same time,with the widespread application of artificial intelligence technology in the Internet of Things,the intelligent interconnection of things puts forward higher requirements for communication.Semantic Communications(SC),as a new generation of communication paradigm,can significantly improve transmission reliability and chain communication.The spectral efficiency of has important application prospects in the future 6G data transmission.As an important part of semantic communication,semantic encoding algorithm is the core technology to realize semantic communication.Existing semantic encoding algorithms are goal-oriented,and the algorithm design takes task distortion as semantic distortion.Starting from the rate-distortion theory,this thesis further proposes the semantic reconstruction distortion,which aims to improve the multi-task generalization ability of the semantic coding algorithm,and consider the adaptively adjustable data compression rate to jointly optimize the rate and semantic reconstruction distortion.The topic selection of the thesis comes from the National Natural Science Foundation of China(No.92067202)and the special fund for basic scientific research business expenses of central universities(No.2021XDA01-1).The dissertation focuses on the image semantic coding algorithm,starting from the rate-distortion theory,extending the traditional pixel-level and feature-level distortion to semantic-level distortion,and studying the image semantic coding algorithm based on semantic reconstruction.In order to adapt to different input sources and channel changes,the image semantic coding algorithm whose compression rate can be adaptively changed is studied.The specific research content of the thesis is as follows:(1)This thesis systematically expounds the research status of image semantic coding algorithms in semantic communication.First,a brief introduction to semantic communication is given,including the development history,basic structure and key technologies of semantic communication,and the importance of current semantic coding algorithms is pointed out.Then,the research status of image semantic coding algorithm and the implementation framework based on deep learning are introduced from the three distortion levels of pixel level,feature level and semantic level,and different design schemes are fully compared.Then it introduces the relevant theoretical guidance of image coding algorithm,and analyzes how to solve the optimization problem of image semantic coding from the theoretical level.Specifically,the theory involved is mainly the rate-distortion theory.As an extension of the rate-distortion theory,the information bottleneck method defines the original reconstruction distortion as the distortion related to the target,which lays the foundation for the semantic distortion.Finally,the key challenges of the current image semantic coding algorithm and the solution ideas of this thesis are introduced,which lays the foundation for the research work of the image semantic coding algorithm based on the rate-distortion theory proposed in this thesis.(2)Aiming at how to enhance the multi-task generalization ability of the semantic coding algorithm,this thesis proposes a multi-task generalization image semantic coding algorithm based on semantic reconstruction.Specifically,a new generalized semantic distortion measure is first established,including reconstruction distortion and task distortion.Based on this,the image semantic coding problem is formulated as a rate-distortion optimization problem.By introducing the Lagrange multiplier βto weight the reconstruction distortion and task distortion,the self-consistent equation of the rate-distortion optimization problem is derived.By solving the self-consistent equation iteratively,the optimal mapping between the source image and the semantically reconstructed image can be determined.Considering the practical application,a semantic coding framework based on semantic reconstruction is constructed by using deep learning technology.Finally,the semantic reconstruction performance of the proposed algorithm is analyzed through simulation experiments,and a variety of traditional image coding algorithms are compared with those based on deep learning.The image compression algorithm shows the performance advantages of the proposed algorithm in terms of reconstruction quality,classification accuracy and target detection accuracy.(3)Aiming at the problem of how to realize the flexible adjustment of the data compression rate of the semantic encoding algorithm,the paper studies the image semantic encoding algorithm based on the adaptive rate.On the basis of the first research point,an entropy model is further introduced to estimate the entropy rate of the extracted semantic information,obtain the entropy rate after expected lossless compression,and realize the joint optimization of rate and semantic reconstruction distortion.Due to the existence of the entropy model,the data compression rate can be adaptively adjusted with the input source.In addition,since bits are reintroduced as the representation of compressed semantic information,this thesis analyzes the effectiveness of the proposed algorithm from the perspective of mutual information with the help of CLUB(Contrastive Log-ratio Upper Bound)mutual information estimation tool.Considering the specific implementation of the algorithm,the thesis uses deep learning technology to construct an image semantic coding framework based on the entropy model.Finally,through simulation experiments,comparing various image coding algorithms,the performance advantages of the proposed algorithm are analyzed in terms of reconstruction quality,classification accuracy and target detection accuracy.