Research On Motion Deblurring From A Single Image

As a key information source, motion blurring image plays a significant role ina widespread applications such as space exploration, traffic surveillance andevidence collection in criminal case. Moreover, in some scene the degraded imagewith motion blur is non-repeatable. Therefore, the research on a single imagerestoration with motion blur is of important significance in academic research andreality application.Image restoration is an intrinsic ill-posed problem, and there is no availabletime-related motion information in a single image, therefore there is great difficultyin obtaining blur parameters and restoring the image from a single image. This paperstudies on motion deblurring methods from a single image.In this paper, we first compare and analysis existing blur parameter identificationbased on linear motion model, and focus on the motion direction identificationmethod based on Radon transform. To overcome the influence of the factorsincluding the bright cross in the spectrum and aspect ratio of the blurring image, wepropose a improved motion direction identification based on the blocking Radontransform. In the method, we introduce the concept subregion stripe density tocalculate the number of the spectrum block, choose the candidate block based onspectrum block selection principle, and introduce the factor aspect ratio to modifythe relation between tilt angle for the dark stripe and motion direction angle.Direction identification experiment demonstrates that the improved direction methodcan effectively obtain the motion angle of the real motion blur image and has betterstability.After blur parameter and preprocessing are completed, we take the classicalLucy-Richard deconvolution algorithm on the original blurring image and present thequality evaluation according to the image quality evaluation index based on edges.In addition, we also present a blind motion deblurring method in the gradientdomain, and study the key techniques and the influence that the parameter settingtakes on final restoration result. The experiment demonstrates that this algorithm can well restore the details in the real image.