Research On Near Lossless Compressed Remote Sensing Image Restoration Based On Deep Learning

JPEG-LS and JPEG2000 are two widely used satellite image compression algorithms.Nowadays,with the development of satellite technology,the number of launched satellites is on the rise,and the image data on the satellite is also gradually increasing.Image compression technology is very important for improving storage utilization and saving channel resources,but its progress needs time.At the same time,deep learning has made a breakthrough in a variety of low-level image tasks in recent years,which provides new ideas and methods for image restoration.In addition to designing a new and more efficient compression algorithm,deep learning can be a new method to restore the compressed remote sensing images so that we can increase the compression ratio further.This paper researches the feasibility and effect of this method.We analyze two algorithms in detail and find the causes of distortion.Based on deep learning,two different network models are proposed for image restoration which compressed by JPEG-LS and JPEG2000.For JPEG-LS compressed image,JPEG-LS compression algorithm adopts near lossless mode.The distortion of compressed image comes from two aspects.One is from the quantization of prediction error in regular mode,and the other mainly comes from the run mode.In run mode,all adjacent pixel values with small difference will be regarded as same,so it produces horizontal strip image distortion and affects the image quality.we propose a wavelet-supervised convolutional neural network（WSCMM）based on autoencoder.And using the wavelet low-frequency sub-band as the multi-resolution supervision because of the spatial correspondence between the wavelet coefficients and the original image.So that WSCNN can focus on restoring the strip distortion.For JPEG2000 compressed image,JPEG2000 compression algorithm uses lossy mode.There are many reasons lead to the image distortion.The main reason is quantization error of each sub-bands after discrete wavelet decomposition,so the image distortion is mainly manifested as blur.The JPEG2000 image recovery network（JRNET）is designed after fully analyzing the algorithm.JRNET correct the quantization error in wavelet sub-bands by making full use of the prior information in compressed data.It can restore wavelet sub-bands and image in discrete wavelet domain and pixel domain respectively.This paper realizes the restoration of JPEG-LS near lossless compression image with NEAR 16 and JPEG2000 lossy compression image with compression ratio 25.Experiments show that the two networks are effective for their respective target images,and achieve great results in both objective and subjective comparison.We prove that the deep learning has a certain resilience to the two compressed images and is a feasible method.