Dynamic Scene Video Deblurring Based On Optical Flow

Motion blur is a common phenomenon of video quality degradation.Motion blur is usually caused by the relative motion between the camera and the shooting scene,such as camera shake or object motion.Motion blur in video often damages valuable information in video and reduces the visual quality of video.Video deblurring can not only improve the visual quality,but also benefit the following video processing such as object detection.However,due to the complexity of the dynamic scene structure and the irregularity of camera motion,the motion blur in video is often non-uniform both spatially and temporally.Moreover,video deblurring is a highly ill-posed problem.Therefore,video deblurring is an important but very challenging task.Video deblurring can be regarded as a highly ill-posed inverse problem.In order to study how to deblur videos,this thesis first analyzes the imaging principle of video blur.For the motion blur that this thesis focuses on,we discuss several common motion models.Based on the formation mechanism of motion blur,this thesis also introduces how to simulate motion blur and synthesize blurry video dataset.For the video deblurring methods,this thesis studies the deconvolutional and deep learning video deblurring methods.Optical flow plays an important role in both methods.For the deconvolutional video deblurring method,it is very important to assume a reasonable blur kernel model according to the motion information.Traditional single image deblurring methods often use homography matrix or affine matrix to approximate the blur kernel of each frame,but these blur kernel approximation methods can only model the motion of the whole image or part of the image,and can not simulate pixel-wise dense motion.In view of the complexity of motion information in video,optical flow is a simple and effective inter-frame motion model,which can be used to approximate motion blur caused by intra-frame motion.In order to deal with motion blur in complex dynamic scenes,a deconvolutional video deblurring method based on piecewise linear blur kernel is studied in this thesis.This method uses bidirectional optical flow to simulate the pixel-wise blur kernel,which can more effectively model the spatially varying blur caused by moving objects.In order to speed up the calculation,a coarse-to-fine multi-scale calculation strategy is adopted.For video deblurring methods based on deep learning,image alignment based on optical flow can enhance the learning ability of the model to the motion information,and effectively utilize the temporal information.This thesis proposes a deep learning video deblurring method based on image alignment and adaptive information fusion.In order to make full use of the complementary information of the adjacent frames in the video,this method performs motion estimation by estimating the optical flow,and aligns the target frame to the reference frame by motion compensation.In order to suppress the artifacts caused by alignment errors,this thesis adopts adaptive information fusion to adaptively adopt the adjacent frame information.In order to fully analyze this video deblurring method,in addition to the traditional methods such as visual contrast and Peak Signal to Noise Ratio(PSNR),this thesis also uses the object detection experiment to verify whether video deblurring as video preprocessing can improve the recognition of objects in video.Experimental results show that the proposed methods can significantly reduce video blur and increase video sharpness.Moreover,the video deblurring method based on deep learning is faster and achieves better visual effect,which can also improve the effect of object detection.