Remote Sensing Image Compression Method And Application Based On Depth Autoencoder

Due to the great mass fervor of artificial intelligence(AI)waged worldwide and the growing maturity of research into technologies such as deep learning,the convolutional neural networks(CNN),the deep learning method obtained through combination and improvement of the convolution computing and the traditional neural network has found increasing applications to different research fields,including image coding.Review of the existing research findings shows that research into the existing image compression mainly focuses on compression of natural images.As a result,study on applications of the CNN to compression of remote sensing images is insufficient.Combining the structural characteristics of the CNN and the binary neural network(BNN),this paper builds the convolutional auto-encoder(CAE)image compression model integrating the dichotomous noise at an attempt to explore new channels to realize compression of remote sensing images combining the deep learning technology.Considering that most image compression models represented by the auto-encoder(AE)are mostly compressed on the basis of natural images,the large-size remote sensing image is chosen as the research object,and the remote sensing images collected from DOTA dataset are adopted as the sample data for experiment.Following that,the image block transformation is used to cut original images of a large size into small-size images to address the high training difficulty in the model realization process,and divide them into the training set and the verification set based on research needs.According to the characteristics of the BNN which can reduce the network complexity,the design mainly includes the CAE compression model containing codes and decoding.The main innovation point of thesis are to add a binary to form a deep convolutional auto-encoder network(DCAE)model for remote sensing image compression.Image processing of the model consists of three stages,including coding,which inputs the original images to be compressed into the coding network.Through the convolution,the spatial scale and number of characteristic patterns of the original images are decreased.Next,image binarization is carried out.The remote sensing image characteristics output by codes are binarized to be translated into occupy space smaller binary coding.After that,the image features are compressed on the basis of coding.The network after binarization can maintain a relatively high compression performance and improve the model's generalization capability.The last step is decoding.Through deconvolution,the image pixels are recovered step by step to obtain the corresponding reconstructed compressed images of input images.In order to strengthen the network stability,thispaper introduces the residual compensation block to the coder and decoder model based on the idea of residual network.After the model is built,the training dataset is used to train the acquired model parameters,and apply images in the remote sensing test set to the model.Then,the images after compression are comparatively analyzed with the traditional JPEG and JPEG2000 images in terms of the compression ratio,peak signal to noise ratio(PSNR),etc.Results show that the compression and reconstruction effect of the CAE designed is superior to that of the traditional image compression.