Research On Image Compression Framework Based On Deep Learning Optimization

With the development of Internet networking,more and more people use pictures and videos to communicate on the network,and they generate huge amounts of data every day,which brings great pressure to transmission and storage devices.Image compression technology has been difficult to meet the compression needs of massive image data by removing the existing information in the image,and there is an urgent need for new and efficient compression technology.Appearance.Deep learning methods using artificial neural networks have shown explosive development in recent years,and have achieved results that exceed traditional methods in many fields including computer vision,speech recognition,and natural language processing.Applying deep learning to solve problems has become a priority for many researchers.In the field of image compression,methods based on deep learning have also made a lot of progress.In terms of reconstruction quality,the coding method based on deep learning has been able to surpass the traditional coding method.Considering such a powerful expression ability of deep learning models,it is very necessary to explore the use of deep learning methods to solve image compression problems.There are two main research directions for image compression based on deep learning: the method of combining deep learning with traditional coding frameworks and the method of completely using deep neural networks.The two directions are explored separately here,the main contributions are as follows: This paper first proposes a compression framework that combines deep learning and traditional coding.Only a neural network is required to pre-process the image input to the traditional compression coding framework to achieve a better ground reconstruction effect.The traditional encoder can use any traditional compression encoding framework.Compared with other methods that use two or more neural networks,the method proposed here does not require special implantation of the decoder used by the user,and is easier to promote.Because the process of encoding and decoding of traditional encoders is not trivial,it cannot be trained by back propagation.In order to allow the neural network to effectively learn the losses caused by traditional coding,we have specially designed a special loss function.As a preliminary experiment,the coding framework shows that our method can effectively improve the quality of image reconstruction.Subsequently,this paper proposes a compression method based on a multi-scale structure using neural networks.In this method,the image is converted into a segmented feature map at the encoding end,and then these feature maps are down-sampled to multiple different scales.At the decoding end,these feature maps of different scales are upsampled to the same scale,through dense connection The dense block extracts the features of the image.Finally,the reconstruction is completed.In the end-to-end joint optimization process,the neural network can obtain image feature information at different resolutions by downsampling the image at different scales,and better reconstruct the image.Different from other methods based on autoencoders,different training bit models need different compression bit rate limitations.Here,the method can choose a single model to achieve compression by selecting the number of downsampling alternatives used by the encoder.Variable bit rate during the process.Finally,an experiment was conducted and compared with several other compression encoding methods.The effectiveness of the compression coding framework proposed in this paper is proved.