Research On Remote Sensing Image Restoration Algorithm For TDI CMOS With Ultra-High Line Frequency

Space optical remote sensing imaging technology has attracted much attention in the fields of military reconnaissance,marine exploration and landform mapping due to its advantages of concealment and efficiency.To obtain higher resolution remote sensing images,our subject developed a space optical remote sensing imaging system based on the TDI CMOS with ultra-high line frequency and high accumulation level.Because TDI CMOS uses ultra-high line frequency for progressive scanning,the exposure time of a single scan is bound to be shortened.Although the high accumulation series plays the role of accumulating enough imaging charges,it cannot reach the theoretical value due to the restriction of phase shift matching accuracy.Therefore,the obtained image has the characteristics of insufficient overall energy and low signal-to-noise ratio.Moreover,the imaging image quality of the system is inevitably affected by factors such as remote sensing imaging equipment and environment,i.e.,the Modulation Transfer Function(MTF)decreases.In the lack of prior knowledge,the MTF compensation technology,which based on relevant theoretical knowledge and mathematical model construction,will face failure,unable to cope with the interference of imaging environment and obtain accurate compensation convolution kernel.The blind image restoration technology is more suitable for the improvement of MTF in the situation.It can estimate the blurry kernel,i.e.,Point Spread Function(PSF),by using the relevant prior knowledge only through the acquired degraded image to eliminate the blur in the image and restore the real image.This paper studies the imaging mechanism of remote sensing imaging system based on TDI CMOS with ultra-high line frequency,and constructs the corresponding MTF compensation model.We designed two blind restoration algorithms for remote sensing images on the local binary pattern(LBP)prior and overlapped patches’ non-linear(OPNL)prior respectively.The experiments show that the algorithms developed in this paper have good convergence and stability,and can be competent for the restoration of remote sensing image.In addition,according to the problem of image evaluation in the process of restoration,this project also proposes an image quality classification system to realize the preliminary discrimination function of image quality,which has certain engineering significance.The research content of this paper mainly includes the following four aspects:Firstly,we have studied the imaging mechanism of the ultra-high line frequency TDI CMOS remote sensing imaging system,and identified the reasons for the degradation of the imaging quality of the system.This paper investigates the traditional MTF compensation method and image restoration method,explores the key and characteristics of both,and finally determines the image quality improvement scheme and constructs the corresponding MTF compensation model.Secondly,a blind restoration algorithm based on LBP is designed for remote sensing images.The algorithm focuses on the similarities between the local binary mode of the clear image and the blurred image.The key pixels of images will be filtered by setting the LBP threshold,and the corresponding processing is carried out for different pixels and their gradients respectively,so as to sharpen the tiny details of the important edges and prevent excessive sharpening at the same time.An effective algorithm is established to recover clear remote sensing images,which based on the projected alternating minimization and semi-quadratic splitting method.Thirdly,in order to obtain more stable processing results and reduce the selection range of relevant parameters,this paper designs a remote sensing image restoration algorithm based on nonlinear prior of overlapping patches.The image is divided into several overlapping patches,and the features are extracted in the unit of patches.The prior is constructed by the ratio of extreme pixels.The relevant solving algorithm is established to restore clear remote sensing images,which based on the projected alternating minimization and semi-quadratic splitting method,combined with fast threshold iteration,sparse mapping matrix and fast Fourier transform.Fourthly,on the premise of image processing,the image may have features such as structure destruction,excessive sharpening,and serious artifacts,which will seriously interfere with the evaluation of the image and cause misjudgment on the evaluation of the algorithm capability.Using the frequency band information to obtain the features that are conducive to distinguish between high-quality and low-quality images,a machine learning model based on decision tree support vector machine(SVM)is constructed,which enables it to progressively distinguish five types of images: images with damaged structure,normal images with excessive sharpening and residual artifacts,dim images with excessive sharpening and residual artifacts,high-quality normal images and high-quality dim images,and the overall accuracy reaches 92.5%.