Restoration Of Remote Sensing Images Based On Point Spread Function Estimation

Despite the increasing demand for high-quality images in environmentalmonitoring, ocean surveying, geomorphological mapping and other fields, theremotely sensed images inevitably suffer from image degradation due to theatmospheric disturbances, camera defocus, satellite vibration, etc. The imagedegradation corrupts the image quality and furthermore affects the image analysisand interpretation. In order to improve the image quality, the degraded imagesrequires restoration. Based on the features of spaceborne optical imagers, therestoration is divided into three sectors in this paper, namely, destriping, image-basedestimation of the point spread function (PSF), and the image deconvolution based onthe estimated PSF.Due to the nonuniformity of Charge Coupled Devices (CCDs), stripe noise iscommon to most pushbroom-type imagers. For the purpose of improving imagequality and avoiding disturbance of posterior analysis and processing, a newdestriping method is proposed. By taking a quasi-homogeneous region as a reference,ideal data of the striped area can be first estimated and then utilized to computecorrection parameters. The cross-track edge information is extracted and processedusing morphological techniques. Experiments carried out on real remote sensing datademonstrate that the proposed method can bring Peck Signal-to-Noise Ratio from32 dB up to48dB. Compared to existing typical algorithms, our destriping method canachieve lower Image Distortion, higher Inverse Coefficient of Variation and PeckSignal-to-Noise Ratio. By reducing stripe noise and reserving useful details, theimage quality is improved effectively.An accurate estimation of the point spread function of the whole imagingsystem is a crucial step in the evaluation of the image quality and the imagerestoration. Based on the classical step-edge method, we made a profound analysisabout the influence of the tilted angle, size, interpolation, contrast, pixel saturationand noise of the step-edge image. The conditional step-edge image is first detectedand then processed to weaken the impact of sampling phase in order to improve theprecision of the measurement. The experiment on real remote sensing panchromaticdata demonstrates the effectivity and credibility of the new method. On the otherhand, a novel estimation method based on Multichannel Blind Deconvolution (MBD)is proposed here to overcome the drawback of the excessive dependence on groundtargets. Multiple sub-images are extracted from the area with uniform localbackground and alternate minimization algorithm is used to implement the blinddeconvolution. It only takes20iterations to achieve a percentage mean square errorof5.9%with two noise-free sub-images. For two sub-images with thesignal-to-noise ratio of45dB, it can get to5.9%. When applied to Wiener filtering ofthe original image, the PSF obtained by this method achieves better results than thefrequently used slanted-edge method (ISO12233).Aimed at the artifact that arises in traditional deconvolutional methods, adeblurring method based on constrained total-variation regularization is proposedafter the PSF is estimated. A total-variation regularization term is introduced to theminimization function and then solved by applying fast gradient projection algorithmto this non-smooth optimization problem. On the inevitable existence of PSFestimation error and noise, the proposed method does not produce significant ringingand noise amplification. Experimental results based on panchromatic remote sensingimages show that it can preserve mean value and meantime, increase Energy of Details from23.9047to35.2431, the Gray Mean Gradient from20.8893to39.0888,the Energy of Laplacian from21.6187to57.7854. The Structural Similarity Indexbetween original and deblurred images is up to0.9813. Both visual effect andevaluation indicators demonstrate that the proposed method can effectively improvethe quality of remote sensing images.