Research On Encrypted Image Compression And Reconstruction Based On Deep Learning

In recent years,the popularity and mature application of cloud computing technology has made it the mainstream for users to use the cloud to transmit and store data.Considering data privacy protection,users encrypt data before transmission,but because of limited computing resources or no benefit interest for compression,users often only encrypt content data at the sending end,without data compression processing.Given the widespread use of images,this would lead to a dramatic increase in encrypted data in the cloud.However,the limited channel bandwidth resources and cloud storage capacity would prompt the cloud to compress the encrypted image data without accessing the encryption key.In order to restore the original image,the receiver needs to decrypt the image and reconstruct it with high quality in order to meet the needs of users for high quality images.Therefore,in the field of large data and cloud computing applications,how to compress encrypted images efficiently and reconstruct them with high quality is one of the problems that need to be studied in depth.The research of this problem can promote the practical application of cloud computing as well as the reversible information hiding in encryption domain.In order to store image data efficiently and reduce the channel pressure of downlink transmission,it is necessary for the cloud server to compress image data.Because the cloud server does not have the encryption key of image data,it simply conducts the uniform downsampling,which remains the spatial distribution of image content.The receiver then performs the decryption to obtain the downsampled image.In order to better reconstruct the original image,it is necessary to make full use of statistical characteristics of the carrier image.Regarding that deep learning well characterizes statistical property of carrier image,it is used in our work to facilitate the image reconstruction,i.e.,the image reconstruction is formulated as a problem of image super-resolution and the deep learning is taken as amethod to solve this problem.As the reconstruction scenario is actually somewhat different to that of the conventional super-resolution,we need to devise the convolutional neural network of deep learning for our situation.In addition,there are also some problems in the existing algorithms for image super-resolution reconstruction using the deep learning,which is summarized as follows: 1)Although the traditional convolution neural network can achieve good reconstruction results,it is limited to the smoother and less detailed images;2)while the deep convolution neural network can fully extract the texture and edge information of the image and fit the higher dimensional feature expression,it would lead to problems of gradient vansishing and characteristic loss;and 3)the single-task loss function defined via output and input variables is usually used in the neural network reconstruction algorithm,but it is easy to fall into over-fitting in the reconstruction process,which in turn cannot result in desirable reconstruction quality when the to-be-reconstructed sub-image is an outlier of training samples.In this thesis,theIn this thesis,the reconstruction problem is represented as the super-resolution reconstruction of decrypted but not de-compressed sub-images.The problem is then solved by designing and optimizing the deep learning network,aiming to reconstruct the original image with high quality.To this end,in the thesis we conduct the following researches,i.e.,(1)On the problem of image feature learning,this thesis proposes a super-resolution reconstruction algorithm of neural network containing residual block structure.Specifically,according to the requirement of sub-image reconstruction,we design a neural network,and further propose the content supplementary structure and Jump-Layer structure to better extract image texture information and improve the issue of feature loss.(2)In this thesis,the multitask loss function is studied for the optimization of reconstructed networks.The multitask loss function is composed of three weighted combinations of single-task functions.In the process of network optimization,the backward propagation algorithm acts on three single-task outputs in parallel.Because three single-task outputs share the underlying hidden layer unit parameters,the feature representation for one task in the hidden layer can also be used by other tasks.The multi-task loss function is used to adjust the parameters,which improves the convergencespeed and generalization ability of the network.(3)Based on the above optimized reconstruction network and loss function,two encryption image compression and reconstruction algorithms based on depth learning are designed in this thesis.Among them,stream cipher is used for encryption,uniform down-sampling is used for compression,and deep learning super-resolution reconstruction network based on residual block structure and its improved version are used for reconstruction.Experimental simulations on authoritative training data sets and test sets show that the proposed schemes encrypted image compression and reconstruction significantly improve the compression efficiency.In addition,the proposed scheme is comparable to other classical algorithms of image super-resolution,which is because the decrypted but compressed sub-image is essentially not the same low resolution image in the conventional super-resolution scenario.In summary,the proposed scheme improves the reconstruction quality and has certain application value in practice.