Research Of High Ratio Remote Sensing Image Compression Technology Based On Deep Learning

In recent years,with the increasing of remote sensing satellites and the resolution of remote sensing images,the data size of remote sensing images is increasing geometrically.The increasing scale remote sensing images makes the data quality of remote sensing applications much higher and the contents more diversified.At the same time,it also increases the difficulty of storing and transmitting remote sensing images.The development of high compression ratio techniques for remote sensing images will help ease the pressure of remote-sensing image storage and transmission.However,while large-rate compression ratio(>=32)tasks are performed by the transform method,problems of visual distortion and high-frequency information lossing and high algorithm complexity of the task of hyperspectral remote image compression are raised.Aiming to solve the above problems,this paper introduces deep learning to the task of remote sensing image compression,and realizes the sparse representation of remote sensing images and the high-dimensional to low-dimensional mapping of remote sensing images by memorizing a large number of remote sensing image data to meet the requirements for high compression ratio of remote sensing images.The main work of this article is as follows:Aiming at the problem of visual distortion,a high ratio compression method of remote sensing image based on sparse stack auto-encoder is designed.The basic idea is as follows: First,a sparse auto-encoder network is built by stacking sparse auto-encoders.Second,training algorithm is introduced by introducing a joint measure of structural similarity term and image distortion term in the loss function.Finally,the output of the hidden layer at the end of the network is quantized and encoded and converted into a binary stream to achieve large rate compression.The experimental results show that the proposed method can improve the SSIM numerical value and the PSNR numerical value of the SPIHT and JPEG methods in high compression ratio tasks,and can obtain better visual effects.Aiming at the problem of high-frequency information lossing,a high ratio compression method of remote sensing image based on lightweight deep convolutional network is designed.The main idea is as follows: First,an 8-layer lightweight deep convolutional network model composed of a coding subnetwork and a decoding subnetwork is designed.Secondly,the binarizer is designed to convert the features of remote sensing images into binary code streams and complete the remote sensing image compression.Experimental results show that compared with the JPEG2000 method,this method can obtain higher PSNR numerical indicators and better subjective visual effects in high compression ratio tasks in the complex scenes such as cities.Aiming at the problem of high algorithm complexity of the task of hyperspectral remote image compression,a high compression ratio method of hyperspectral remote sensing image based on tensor convolution deep neural network is designed.The main idea is as follows: First,design the tensor convolutional deep network of hyperspectral remote sensing images.Secondly,the joint measurement of spectral angle measurement item and image fidelity term is introduced into the loss function for training.Finally,the output of the encoder is quantized and entropy-encoded to achieve large-magnification compression of hyperspectral remote sensing images.Experimental results show that this method can exploit the spatial spectrum features of hyperspectral remote sensing images.Compared with JPEG and SPIHT compression methods,it can obtain higher SSIM numerical indicators and PSNR numerical indicators as well as better subjective visual effects.