ALOS-PRISM遥感影像超分辨率重建

介绍了日本ALOS卫星PRISM三线阵传感器的成像原理和方法,提出了利用PRISM三线阵影像进行超分辨率重建来提高PRISM影像的空间分辨率.提出了新的光流配准算法,该算法将标准互相关配准算法引入到Lueas-Kanade光流配准算法中,大大的减少了误配率,能够有效的消除PRISM Level 1级别的影像之间由于地形起伏所引起的变形.同时,改进了影像的高斯退化模型,在超分辨率算法中,引入了可变退化函数,通过交替最小化(AM)算法对可变退化函数进行盲估计,实验结果表明,超分辨率重建影像与插值影像相比,细节清晰很多,有效的提高了影像的分辨率.实验结果说明了本文配准算法可以达到超分辨率重建的亚像素的精度要求,可以应用于航空遥感影像的高精度匹配,同时也说明了将航空遥感影像的退化函数算子分为高斯退化算子和可变退化算子的思想是正确的,符合实际情况.

Super-resolution reconstruction of ALOS-PRISM remote sensing images

We introduce the Advanced Land Observing Satellite and its Panchromatic Remote-sensing Instrument for StereoMapping (PRSIM)and use the super-resolution reconstruction approach to improve the resolution of thePRISM images. PRISM is apanchromatic radiometerwith2.5meter spatial resolution. PRISM instrumentbelongs to the class ofpush broom sensorand data areacquired by a linearCCDs array. PRISM product are processed into CEOS format for level1B1,1B2R, and1B2G.The image ofLevel1B2G is geometrically corrected data. The PRISM sensor can capture three images in the direction of looking forwards,downwards and backwards from the aircraft or satellite at same time. So we can obtain three images ofLevel1B2G in the samescene. Super-resolution technique can obtain a high-resolution image from observed multiple low-resolution images. The majoradvantage of the super-resolution approach is that itmay cost less and the existing low-resolution imaging systems can still beutilized. There is a greatneed to have fine spatial resolution datawith high fidelity and consistence in geo-referencing and intensity(tone)in the studies of land coverand land use, and their changes. In view of this, we present amaximum a posteriori estimationframework to obtain a high-resolution image from the PRSIM images ofLevel1B2G. This super-resolutionmethod is composed oftwomain steps. In the first step, we presenta hybrid optical flow registrationmethod to dealwith the deformationwhich is broughtby hypsography. In order to improve the registration accuracy of PRISM Level1B2G Images, we propose a new optical flowregistrationmethod. This approach uses theNormalizedCross-Correlation registration algorithm beforewe useLucas-Kanade opticalflow registration algorithm. Optical flow is the distribution of apparent velocities ofmovement of brightness patterns in an image.Optical flow can arise from relativemotion of objects and the viewer. The Lucas-Kanade registration approach divided the originalimage into smaller sections and assumes a constantvelocity in each section. Then itperforms aweighted least-square fitof the opticalflow constraintequation. It can detectmost local distortions of PRISM image in sub-pixel accuracy, but thismethod may lead tosomemisregister. The Normalized Cross-Correlation registration algorithm can reduce the misregister. So, we take the NCCregistrationmethod to perform coarse registration firstly. The mixture registration method can remove the deformation which isbroughtby hypsography in a greatmeasure. In this second step, to reconstruct the high-resolution image, we apply an iterativescheme based on alternative minimization to estimate the blur and HR image progressively. It is the combination of the bluridentification and high resolution image reconstruction.We also improve the Gaussian PSF assumption mode,l and introduce thevolatile blurs into the PSFmode.l By AlternatingM inimization (AM)algorithm, we can estimate the volatile blurs. Image qualityassessmentplays an important role in image super-resolution reconstruction. Peak Signal-to-NoiseRatio (PSNR)andMean SquaredError (MSE) are the most widely used objective image quality indexes. The two indexes are Full-Reference image qualityassessment. Unfortunately, we can notobtain the originalhigh resolution image in the super-resolution reconstruction process. Sowepropose two noreference image quality assessmentswhich are entropy andMean Grads. Experimental results show that our super-resolution method is effective in performing blind SR image reconstruction with PRISM images and our super-resolutionreconstruction algorithm has better performance in edge preserving than bicubic interpretation. The resolution of PRISM image isenhanced effectively. The enhancement show that themixture registrationmethod can reach sub-pixelprecise and themodification oftheGaussian PSF assumptionmodelcorrespond to the actualPSF ofPRISM images. TheAM blind super-resolution approach can beused to enhance the resolution ofaerialand remotely sensedimages.