The Processing Algorithm Study For Remote Sensing Images Based On Multi-scale Geometric Transformation

Remote sensing image plays an irreplaceable role in national defense and civil applications. Its imaging mechanism is different from that of visible images, so the study of remote sensing image processing using its characteristics is particularly important. Multi-scale geometric analysis has10years of history since proposed. The method has a good time-frequency performance and gives a good solution to deal with problems corresponding to multi-dimensional signal processing. However, multi-scale geometric transformations have many disadvantages that should to be improved. For example, some multi-scale geometric transformations lack shift invariance and analysis, and also have less selective direction. What’s more, people will get different effects from different multi-scale geometric transformations for different characteristics of remote sensing images. The author mainly focuses on the research of multi-scale geometric transformations and its applications in remote image processing.This dissertation mainly focuses on widely used multi-scale geometric transformations such as Contourlet, dual tree complex wavelet (DTCWT), Shearlet and hyperanalysis wavelet (HWT). Their main application area is the radar image processing, including SAR image denoising, edge detection, image separation, the airport foreign object debris detection and remote sensing image fusion.This dissertation studies the following two aspects on the basis of our predecessors. On one hand, we improve the property of the existing multi-scale geometric transformation and create new multi-scale geometric transformation with better property. On the other hand, different multi-scale geometric transformations and models are constructed depending on the application scenario in the field of remote sensing image processing. In the course of the study, we want to construct some better multi-scale geometric transformations and propse several methods suitable for remote sensing image processing. The main contribution of this dissertation is as follows.1. SAR image de-noising based on multi-scale geometric transformationSince the existing multi-scale geometric transformations lack direction selection and shift invariance, this article improved several multi-scale geometric transformations with shift invariant features, including wavelet-Contourlet transform, the complex Contourlet transform, the local hybrid filter, complex Shearlet transform and shift-invariant two-dimensional hybrid transform. These improved transformations overcome the above disadvantages, enrich multi-scale geometric transformations, and they are also easier for radar image denoising.Several denoising models based on the multi-scale geometric transformation based on improved Contourlet and Shearlet have been proposed for the imaging features of SAR image, including the cycle spinning denoising model based on Wavelet-Contourlet transform, Gaussian mixture denoising model based on the complex Contourlet transform, denoising model based on local hybrid filter, bivariate denoising model based on Shearlet, denoising model based on sparse representation, Bayesian shrinkage denoising model based on sparse representation, Gaussian mixture denoising model based on the complex Shearlet and de-noising model based on2-D shift-invariant hybrid transform for the airport radar image.Simulation results show the validity and reliability of the proposed denoising model. In view of the above algorithms, this dissertation summarizes the common framework of denoising method based on multi-scale geometric transformation, and gives the algorithms’of comparisons, which analyze their advantages and disadvantages in radar image denoising in order to facilitate further research in the future.2. Multi-scale geometric transformation for SAR image edge detectionSince the existing multi-scale geometric transformations in SAR image edge detection do not take full advantage of the directions information of multi-scale transform and the fusion rules of multi-scale edge information are relatively simple, this dissertation summarizes the typical steps of SAR image edge detection based on multi-scale geometric transformation, and three SAR image edge detection models based on multi-scale geometric transformation are proposed.The first edge detection algorithm is based on local hybrid filter denoising model. We have improved several steps of the SAR image edge detection algorithm based on multi-scale geometric transformation. First, smoothing process is improved. Then the Canny operator based on ROEWA model is used for single scale edge detection. Finally, the edge of the scale information fusion using evidence theory is improved as well.The second edge detection algorithm is based on sparse denoising and least squares support vector machine. Firstly, use sparse representation to de-noise, and then use the least squares support vector machine to detect edge.The third edge detection algorithm is based on the sparse representation denoising model. Sparse representation denoising is an iterative denoising model. Use morphological operators to detect the direction edge information during each iteration and fuse all information to complete the whole edge using evidence theory.Finally, the proposed algorithms are compared, and their advantages and disadvantages in the SAR image edge detection are analyzed in order to facilitate further research in the future.3. Multi-scale geometric transformation for image separationIn the study of stars trajectory, the points and curves in astronomical images need to be separated, but the complexity of existing algorithms is very high and their computation time is too long. So in this dissertation, three new image geometric separation dictionaries for image separation are proposed. One is based on complex Shearlet and biorthogonal wavelet dictionary, another is based on circular symmetric Shearlet and biorthogonal wavelet dictionary, the last one is based on the hyperanalysis Shearlet and biorthogonal wavelet dictionary. This article also uses a new iterative algorithm for image geometric separating in the last dictionary. To get the objective evaluation of image geometry separation efficiency, this dissertation presents an evaluation of criteria separation. Experimental results show the effectiveness of the proposed algorithm. The dissertation also analyzes the advantages and disadvantages of the algorithm, in order to facilitate further research.4. Multi-scale geometric transformation for remote sensing fusionSince the fusion effects of the current image fusion rules in transform domain are relatively poor, which are caused by the artificial texture, the dissertation proposes a new fusion method based on Shearlet transform with guided filtering. This method greatly suppresses the artificial texture by using the spatial continuity of the image. The experimental results show that the algorithm not only effectively improves the image fusion visual effects, but also has good robustness, can be applied to the image fusions multi-gathering imagery and remote sensing images, as well as other types of image fusion.