Research On Image-based Image Motion Self-matching Digital Domain TDI

Time delay and integration（TDI）is an important imaging mode for satellite-borne or airborne cameras to achieve high resolution images under the condition of fast relative speed.Recently,many work have done in different aspects: From TDI realized in charge domain in charge-coupled device（CCD）,to TDI implemented in analog domain in additional circuits of complementary metal oxide semiconductor（CMOS）,and then to realize the TDI in the digital domain with the help of field programmable gate array（FPGA）.It can be seen that it has become a new trend to get rid of the limitations of detectors and improve the flexibility to expand the application range.To further increase the flexibility of digital-domain TDI,this dissertation proposes an image registration based TDI framework,the basic idea of which is to combine the flexible digital domain TDI with image-based image motion compensation to realize image motion self-registered TDI,thereby reducing the requirement of image motion compensation accuracy for high-resolution TDI imaging,and further expanding the application range of TDI.This dissertation first conceives a framework of image registration based TDI,during which a basic technical route is determined: firstly implement a denoising algorithm suitable for dim-light images,reduce the influence of noise on subsequent registration,and then implement an efficient registration algorithm,and finally verify the validity of the overall algorithm.To denoise low-light images,the noise characteristics of dim images is first modeled and then to analyze the characteristics of dim remote sensing images,according which the BM3D-GT＆AD is proposed to denoise a series of short-captured low-light images.This algorithm exploit the difference between the regularity of detail distribution and the randomness of noise distribution in low-light remote sensing images,then to introduce the Gaussian-distributed thresholds in hard thresholding;Considering the interference of large noise on the statistics in small image blocks,in order to improve the rationality in 3D-block grouping,normalized angular distance is applied to replace the regular similarity measure.The experiments show that the proposed BM3D-GT＆AD has incremental peak signal-to-noise ratio（PSNR）compared to the original BM3 D,but show the best structure similarity（SSIM）compared to other state-of-art denoising algorithms.From the image denoising effect,BM3D-GT＆AD retains more details while denoising.This can significantly reduce the registration error,and improve the stability.In image registration process,considering the influence of remaining noise and the real-time requirement,an improved Fourier-Mellin transform（FMT）based image registration method was first proposed.When estimating the relative shift,it first calculate weighted column standard deviation of FMT spectrum of images,where the weight is mainly assigned to the intermediate-frequency components.According to the weighted column standard deviation,a special filter is designed to estimate the relative shift by exploiting the intermediate-frequency components.By this way,the robustness to noise is improved from 25 to 40.To improve the registration accuracy,a descriptor that utilize inter-feature relative azimuth and distance（IFRAD）is also proposed.It first apply simple feature detection method like FAST to obtain some candidate feature-points,then apply regulation to select secondary features and to determine primary features.Finally the azimuth and distance from each primary feature to all other secondary features are calculated to obtain orientation-intensity histogram（OIH）,which is used to quantitively describe each primary feature.The scale-invariance of OIH is guaranteed by adopting cosine distance in feature matching process.Compared with other feature-based registration algorithms,it eliminates the time-consuming process such as calculating local gradient changes and building image pyramids.To verify the effectiveness of the overall algorithm,two types of experiments are carried out:（a）computer simulations and（b）outfield push-broom experiments.In the computer simulations,two evaluation indicators are firstly introduced:（a）the algorithm’s tolerance to image dimming and（b）the image registration accuracy,and then the pros and cons of the two registration algorithms and their applicable scope are compared,and the validity and necessity of BM3D-GT＆AD is also verified.In the outfield push-broom experiments,the validity of the image registration based TDI is verified from the image details.In comparison to current technology of TDI-CCD or TDI-CMOS,the technology proposed in this dissertation improved the flexibility of TDI-CMOS,for example,the integration stages can be continuously adjusted,the direction of push-broom is arbitrary,reduced complexity of structures,etc,which meets the development needs of miniaturization.