Research On Gaussian Blurred Images Restoration And Matching

Image matching plays a great role in visual navigation systems,which aims to get the precise position of the real-time image in the reference image while the real-time image is captured in the same scene with the reference image but under different conditions.However,the captured real-time image is usually inevitably degraded since the real environment is complicated,such as Gaussian blur,which will reduce the image matching accuracy severely.And most of the existing matching algorithms don't consider the degradation phenomenon.Therefore,it is of great significance to study the blurred real-time image matching,whose accuracy will benefit from it a lot.In this thesis,we carry out research on the image matching of Gaussian blurred real-time image.And we propose to start from the two aspects of Gaussian blurred image restoration and image matching,exploring the relationship between restoration and matching.In the end,we propose three methods to achieve good results for both image restoration and matching.Since the degraded image affects the matching accuracy seriously,we propose a two-phase restoration and matching method in which we first reconstruct the Gaussian blurred image with the proposed deep generative adversarial network(GAN)and then we could use the reconstructed sharp image to get its position with the method of sparse representation.With this method,we can achieve good results for both tasks.But in the real world,the captured scene is usually complicated with many repeated structures,which puts more demands on the quality of the restored image.Therefore,we further introduce the grayscale image of edge extracted from the blurred image and propose another two-phase method that uses the introduced edge priori to restore the blurred image and further applies the restored image to matching task.And in order to get the trained model,we need put the blurred image and the grayscale image of edge along with the ground-truth of the blurred image into the deep edge network.The results of image restoration and matching improve more.In the above two-phase methods,the image restoration and matching task are performed separately and do not implement a multi-task integrated output.At the same time,the matching results do not promote the recovery task.To solve these problems,we propose to use the matching result as a priori information to promote the restored image quality because a higher quality restored image can further promote the matching accuracy.The two can be iteratively promoted together.In order to achieve more accurate matching results to help restoration,we consider both local and sparse information and adopt distance-weighted sparse representation to obtain better representation coefficients.By iterative restoring the input image in pursuit of the sparest representation,our approach can achieve restoration and matching simultaneous,and these two tasks can benefit greatly from each other.In this thesis,we propose three robust blurred image matching methods,which have important practical significance since they achieve both high-precision matching results and restored image of good quality.