Research On Super-Resolution Reconstruction Technology In Frequency- And Spatial-Domain For Remote Sensing Images

Image resolution is one of crucial indices of image quality. Remote sensing image is an important kind of images, whose resolution has vital effect on following target interpretation and recognition. The super-resolution reconstruction of images means the techniques of reconstructing one or more frames HRIs (high-resolution images) from one or more LRIs (low-resolution images) that include the same scenes.Aiming at the characteristics and super-resolution demands of remote sensing images, this paper carries through the innovation researches for the super-resolution reconstruction technologies of images in frequency domain and spatial domain under unchanging the hardwares of imaging system.In the aspects of super-resolution reconstruction technologies in frequency domain for remote sensing images, improved single-frame super-resolution algorithm and gaussian re-sampling function are used to break the two restriction conditions: the input frame number which is the least 4 frames and the sub-pixel shifts between frames which must contents with the determinate relations, then the improved dealiasing method in frequency domain under inputing two frames is established. Firstly, spectrum spread and compensation based single-frame super-resolution algorithm is improved. Spectrum transform and enhancement filter is proposed, which will be performed when ring degree is trivial, so computation quantity is decreased largely under not decreasing super-resolution effects; The self-adpative calculation method of the control paremeter-P value that denotes ring degree is proposed by using quintic polynomial fitting, which can estimate ring degree more veraciously and conveniently. In the improved dealiasing method in frequency domain with inputing two frames, firstly, single-frame super-resolution algorithm with fourfold-magnified model along row and column for two inputed LRIs are performed respectively, and then, Gaussian re-sampling function is used to perform 1/4 down-sampling along row and colum of the result images of single-frame super-resolution respectively, so 32 frames with the same resolution such as inputing LRI are obtained, therefore the restriction condition to frame nomber is broken; The selection method of 32 LRIs is studied and 12 frames meeting shifts constraint conditions are selsected to implement dealiasing algorithm in frequency domain, so the constraint condition to sub-pixel shifts between frames is broken. A series of experiment results indicate that the proposed algorithm is very effective. PSNR of result images of simulation remote sensing experiments can be increased 6~8dB. The contrast improvements of result images achieve 11~12dB for real two frames input. The resolution of result images processed for input remote sensing image with 2.0m and 3.0m resolutions can be enhanced to above 1.54 times and 1.77 times.In the aspects of spatial domain super-resolution reconstruction technologies of remote sensing images, the PMAP(Poisson Maximum A-posteriori Probability) estimating technology,the POCS(Projection Onto Convex Sets) estimating technology and the PMAP/POCS fusion optimization super-resolution reconstruction technology are innovatively studied in turn so as to obtain better super-resolution effect. According to the characteristic that probability distributing of remote sensing image obey poisson distribution, PMAP estimating technology is studied and improved. By deeply analyzing the shortcoming and limited conditions of conventional PMAP estimation method, we introduced three improvements: the down-sampling and shift operations are added in so as to be used in more general imaging model and improved super-resolution effect simultaneously; the "ill-posed" problem is translated into "well-posed" problem by absorbing Tukey regularization term; the pixels that do not satisfy robustness demand is excluded by selecting pixel by pixel and the algorithm robustness is enhanced. Therefore,the robust generalized PMAP(RGPMAP) algorithm is founded. The experiment results show that the proposed RGPMAP algorithm has good super-resolution effect and strong robustness. The PSNR of result image can be inproved above 5~8dB and 8~10dB for 2/4 frames input with 0.4 pixel registration error or 25dB Gaussian noise. Their effect on RGPMAP algorithm result is below 0.7dB.In the same way, aiming at the weak robustness of conventional POCS estimating technology, a threshold of the difference of corresponding pixels between LR input image and degradation image of HRI is added to carry through the operations of selecting pixel by pixel, so that the POCS algorithm is performed only for those pixels which satisfy robustness condition. Thereby, the robust POCS (RPOCS) super-resolution reconstruction algorithm is established. The experimental results indicate that the proposed RPOCS algorithm has the good performances which have not only strong robustness but also better super-resolution effect. The PSNR of result image can be inproved above 6~9dB and 12dB for 2/4 frames input with 0.4 pixel registration error or 25dB Gaussian noise. Their effect on RPOCS algorithm result is below 0.8dB.Because the set theory based POCS method has the merits of strong absorbing prior information and the probability statistical theory based PMAP method has the strong ability of recovering high-frequency information, the proposed RGPMAP method and RPOCS method are further fused in order to fuse the merits of the two kinds of methods. Iteration order and iteration times of the two kinds of methods are studied and then RGPMAP-RPOCS fusion super-resolution algorithm is founded. The experiments show that the super-resolution effect of the proposed fusion algorithm surpass one of RGPMAP and RPOCS. PSNR of fusion algorithm results are improved 2~3dB to compare with RGPMAP algorithm and RPOCS algorithm in simulated remote sensing image experiments. The contrast improvement of result images achieves more than 13dB in the experiments of input two frames of real remote sensing images. The resolution of result images processed for input remote sensing image with 2.0m and 3.0m resolutions can be enhanced to about 1.75 times and 1.90 times.