Research On Deep Auto-encoder Based Image Lossy Compression

With the popularization of smart terminal equipment,a large number of image and video contents are generated every day.As a key technology for data storage and transmission,image coding has always been the focus of people's research.In the last decades,the development of traditional image coding has gradually encountered bottlenecks.On the one hand,due to the continuous increase in the calculation and storage complexity of various modules in the coding standard,it will be more difficult to improve the performance according to the current coding mode in the future;on the other hand,with regard to the rapid development of intelligent applications such as classification and detection,existing coding algorithms cannot cope with the needs.To solve the above problems,deep learning-based coding algorithms provides new solutions.Considering that auto-encoder is the basic structure of the deep learning coding framework,this thesis focuses on the study of image lossy compression algorithms based on deep auto-encoders.The main contributions of the thesis are summarized as follows:(1)A multi-scale auto-encoder image compression algorithm with adaptive bits allocation is proposed to solve the problem that traditional auto-encoder networks do not make full use of the features of each layer and cannot perform adaptive bit allocation.The algorithm combines the features of different scales in the encoder,processes the features by using the adaptive bit allocation module,and then transmits them to the decoder.The experimental results show that the proposed algorithm has obvious advantages over JPEG in subjective and objective quality,and saves 3.71% bitrate compared with JPEG2000 in MS-SSIM evaluation.(2)An image compression algorithm for maintaining classification accuracy is proposed to solve the situation that the classification accuracy of the image is reduced after the compression algorithm.The algorithm introduces classification feature loss on the basis of rate-distortion loss,and experimentally determines the multi-scale feature fusion scheme and loss function fusion weight,so that the network can optimize the ratedistortion performance while maintaining image classification accuracy as much as possible.The experimental results show that the proposed algorithm can achieve the objective quality similar to BPG under the MS-SSIM evaluation,and the reconstructed image is 2% higher than BPG in the accuracy of Top-1.(3)A region of interest(ROI)enhancement image compression algorithm is proposed to solve the problem of the quality degradation of the ROI in the compression process.In the algorithm,the ROI feature enhanced encoder is introduced to enhance the ROI in the original image features,and the feature of the ROI is fused with the original image feature to enhance the region of interest by calculating the importance scores.The experimental results show that compared with JPEG,JPEG2000,BPG and some deep learning compression algorithms,the proposed algorithm has better objective quality in reconstructed ROI and whole image.Using MS-SSIM evaluation,in the case of the same compressed image quality,the proposed method saves 18.71% bitrate compared with BPG in the entire image,and save 38.64% bitrate compared with BPG in the ROI.