多时相遥感图像复原方法的研究

Research on Multi-temporal Remote Sensing Image Restoration

为了充分利用卫星所获得的对同一地区多次观测图像来进行清晰化,选择了多时相遥感图像复原技术,利用各观测图像间相似但又有差异的信息,重建出反降晰图像以获得更多的细节。论文建立了多时相遥感图像复原的数学模型,估计出观测图像与参考图像之间的相对运动,选择约柬最小二乘法对解进行规整化,获得复原结果。运用该算法对卫星拍摄的多时相遥感图像进行处理,并选择三种衡量标准对复原结果进行定量分析,验证了算法的有效性和实用性。

In practice,images of high resolution are necessary and preferred.The most direct solution to increase resolu- tion is to improve that of imaging sensor.However,the solution may not be feasible due to the growing cost and limitations on current image sensor and optics manufacturing technology. In recent years,many attentions have been attached on a technique called'super-resolution(SR)'which provides an alternative for increasing the resolution of the acquired images.The super-resolution reconstruction problem refers to restoring a high resolution image from multiple low resolution images degraded by warping,blurring,noise and aliasing. The core idea of SR is that the observed low resolution images contain slightly different views of the same object.In this case,the requirement of total information about the object is much higher than information in each frame.If the object doesn't move and is identical in all frames,no extra information can be extracted from the low resolution images. The SR algorithm can be divided by their domain:frequency and spatial.Even since Tsai and Huang(1984)dem- onstrated how to get super-resolution in frequency domain,much work has been devoted to this problem.Although the fre- quency domain methods are intuitively simple and computationally cheap,they are sensitive to model errors,and only pure translational motion can be treated,so current researches are mostly concentrated on spatial domain in which allows more complicated motion models and prior knowledge can be taken into account to improve the performance of reconstruction. To analyze the SR reconstruction problem comprehensively,it's necessary to formulate an observation model that re- lates the original high resolution image to the observed low resolution images first.Several observation models have been proposed in previous literatures.According to existing models,SR reconstruction is related to motion estimation,image restoration and interpolation.There are relative motions among each observed images,and we have to estimate the motions to align each low resolution image to a reference image before we can accumulate information from the observed images. After that,image restoration should be taken because the low resolution images are blurred in the formation model.It's an ill-pose inverse problem which doesn't have direct solution and usually requires regularization(applying some constraints according to prior knowledge). Because SR reconstruction can overcome the limitations of imaging system to improve image resolution under some conditions,it's become more attractive,especially for the situations in which it's easy to capture multiple low resolution images.Remote sensing imaging is one of such situations.Along with the speedup of remote sensing technology,it's much convenient to get multiple images of the same place,but these images usually are not satisfied for our need of high resolu- tion.In this paper,we are trying to reveal more details from the observed low resolution images. The image formation model is introduced first.And then a motion estimation algorithm based on 6-parameters affine transformation is proposed.Finally constrained least squares method is chosen to regularize this ill-posed inverse problem. Practically acquired remote sensing images are used in the experiment,and three criterions are selected to evaluate the quality of restored image,which demonstrate the efficiency and practicability of the algorithm.