Remote Sensing Compression Based On Generative Adversarial Network

With the development of earth observation and remote sensing imaging technology,the number,width and resolution of remote sensing images are increasing substantially.The continuous emergence of a large number of high-resolution images has increased the pressure on the satellite-to-ground transmission system,so it is imperative to develop efficient and high-ratio remote sensing image compression methods.Compared with natural images,remote sensing images have the characteristics of a wide variety of ground objects with rich details and textures.When using traditional compression methods,there are problems such as serious detail distortion and high algorithm complexity.Deep learning developed in recent years has excellent feature extraction capabilities and a highly parallel mechanism,which can quickly map space-time-spectrum high-dimensional remote sensing images to a lower feature space.All the above methods have achieved good results.However,remote sensing image compression based on deep learning still have some problems:(1)The compression model is highly dependent on training data.When the distribution of the training dataset is consistent with the test dataset,the model can achieve good results;and when the distribution of the train dataset is inconsistent with the test dataset,the model performance is poor.(2)The compression ratio is limited by the network structure,and it is not easy to adjust.Since the deep learning model has a fixed network structure,it can only reduce the image to a fixed size code stream,and the adjustment of the compression ratio needs to be realized by retraining the model.(3)When compressing at high ratio,it is easy to lose detailed information such as the edge contour in the remote sensing image.The emergence of generative adversarial network brings a new learning method for deep learning.It models the distribution of images with the idea of game,rather than just learning the characteristics of training images.The deep learning model needs to store the prior information required for image decompression in the network parameters.The higher the compression ratio,the more prior information required for image decompression.When the compression model is trained,the parameters learned by applying the pixel-by-pixel loss model fluctuate greatly with different images,and the generation of the adversarial network maps the original image and decompressed image to the feature level,which reduces the large fluctuation of the parameters and helps the model learn more a priori information.Therefore,compared with the traditional deep learning image compression methods based on convolution neural network and recurrent neural network,the image compression method based on generative adversarial network has advantages of high ratio and great reconstruction quality,which is expected to overcome the defects of poor generalization performance of the traditional deep learning model.Therefore,the dissertation studies the remote sensing image compression based on the generative adversarial network.The main work contents are as follows:1.Aiming at the poor generalization performance of image compression model based on deep learning,a remote sensing image compression method based on sparse flow adversarial model(SFAM)is proposed.Firstly,SFAM can carry out accurate latent variable reasoning by designing a series of reversible modules to fit the image distribution,however,the original flow model does not have sparsity.Therefore,a reversible SFAM is designed.SFAM uses the flow model framework to learn a reversible and stable mapping.The mapping mines the sparsity of the input image by combining linear and nonlinear transformations,and reduces the dependence of the model on a specific dataset by constraining the reversibility of the model structure.Secondly,multi-scale transformation is introduced into the SFAM,which increases the sparsity of linear transformation and reduces the bit stream generated in the compression process.Finally,sparsity adversarial loss is introduced into the training of SFAM to generate more sparse features and achieve efficient compression.Experiments show that when the natural image dataset ImageNet and remote sensing image dataset VPVR with different resolutions are trained at the same time,the PSNR of this method in the range of 4 to 64 times compression ratio is 4%～12%higher than that of JPEG2000.2.Aiming at the problems of artifacts and serious loss of details in the current image compression method based on deep learning,a high ratio image compression method based on detail fidelity adversarial network is proposed.Firstly,the existing generative adversarial network only has global constraints on the image distribution in training,and the quantization operation will lose a lot of gradient information in end-to-end training.Therefore,this dissertation constructs a detail fidelity generative adversarial network to increase the local constraints of the adversarial network on the generator.Thus,the artifact and detail loss caused by the model of adversarial network training with only global constraints are reduced.Secondly,in order to reduce the loss of network gradient back propagation caused by quantization operation,the quantization loss term is introduced in this method.The quantization loss term reduces the information loss caused by quantization operation by increasing the consistency of the gradient before and after quantization operation during training,that is,minimizing the difference of the gradient before and after quantization during back propagation.Experiments show that on different ground feature types datasets and remote sensing image datasets VPVR with different resolutions collected by GF2 satellite,the PSNR index of this method is 1%～11%higher than that of JPEG2000 when the compression ratio is 256.3.According to the characteristics of rich texture and dense information of remote sensing image,inspired by the symmetrical structure in classical compression methods,a symmetrical lattice generative adversarial network(SLGAN)for high-ratio compression of remote sensing image is proposed.In this method,a generator is constructed by using the same number of symmetrical coding lattices and decoding lattices.The coding lattice is responsible for generating the code stream to be compressed,and the decoding lattice is responsible for decompressing the code stream to be compressed into an image.In order to constrain the symmetric codec to learn the symmetric parameters,a discriminator is constructed for each pair of coding lattice and decoding lattice.Symmetric parameters constitute an approximately reversible codec,which can better reveal the image distribution and achieve efficient image restoration.In order to solve this problem,a cooperative learning algorithm is proposed to train symmetric lattice pairs in the generator.In addition,in order to enhance the edge,contour and texture of the decompressed remote sensing image,an enhanced Gaussian Laplace loss is designed as a regular term to train SLGAN.The experimental results on remote sensing image dataset show that the PSNR of this method is 2%一 10%higher than that of the benchmark method.4.Aiming at the problem that the current image compression methods based on deep learning are difficult to control the bit rate,a high-ratio remote sensing image compression method based on residual compensation is proposed.Firstly,a compensation network is designed,which can be combined with any existing codec structure.The fuzzy image restoration problem obtained by compression under the condition of high ratio is modeled as an optimization model of progressive compensation with residual.Secondly,a compensation method with variable ratio is designed based on the size of residual and the grade of quality.The purpose of variable ratio is achieved by separating the residual from the coding transmission of the coding result.Finally,according to the characteristics of dense texture of remote sensing image,a multi-level compensation network is designed,which combines the two network structures of detail compensation and residual compensation,so that the model can fully extract the detail information from the residual.The experiments show that the PSNR of this method is 3%～5%higher than that of the image compression method based on SLGAN when the compression ratio is 256.5.Aiming at the problems of insufficient utilization of spatial correlation and lack of correlation between model components in the existing prediction based hyperspectral image compression methods,a hyperspectral image compression method based on LO norm iterative optimization is proposed.Firstly,an improved super-pixel segmentation algorithm is introduced to realize the initial clustering of spectral lines by combining the spectral angle and Euclidean distance to measure the spectral correlation in the spatial dimension.Then,a linear predictor is calculated in each class of spectral line for prediction,and the residual image between the original image and the predicted image is obtained.By comparing the size of LO norm of each spectral line in the residual image between classes,the class of predictor with the smallest LO norm of the spectral line is selected,and the cluster center and predictor coefficient are recalculated.In this way,the clustering results and predictor coefficients are iteratively optimized.The iterative process optimizes the mutually restricted clustering results and predictor coefficients at the same time.Therefore,this method can obtain less bit stream and more efficient lossless entropy coding results.Experiments on hyperspectral data show that compared with other methods,this method achieves higher compression ratio for corrected and uncorrected images.