Motion Deblurring From A Single Image

Motion blur is a common phenomenon which caused by the relative motion of the camera and the scene or the objects during the capture process. Because of the degraded of the image, we cannot further explore the world, obtaining more information, promoting the development of science. This paper present an algorithm for removing the motion blur from a single image. The algorithm only needs a blurring image but the motion information and the more images, so it simplifies the deblurring process.This paper presents the image blurring model and deblurring model to found a new image deblurring model to a single image. The model decomposed the image deblurring process into two steps: firstly we estimate the motion kernel, secondly we use the estimated image kernel to deblur the blurred image. Then we combine the two steps and constraint both of them each other in order to restore the original image.After that, this paper presents an analysis of the theory and implement to estimate motion kernel based on the motion path. The optical flow is introduced too. And then we introduce the Tikhonov regularization term, spare regularization term and continuity regularization term of motion kernel from the property of kernel. We describe the efficient optimization scheme to calculate the expression which includesthese terms. Then, we describe the multi-scale approach to avoid local minima.And then, we present the common deblurring algorithm such as RL and Wiener. We proposed three improved algorithms: The first one is a Method to reduce the impact of the ring artifacts under the condition of the inaccurate blurring kernel, which assumes that the noise in measurement process is meetting the Gaussian distribution to fit the natural noise distribution. In order to control the non-uniform noise it supposes the first order and the second order of derivatives to satisfy the independent Gaussian distributions. The second one is Real-time Motion Deblurring Algorithm with Robust Noise, which demonstrates the necessity of adding the nature image gradient constraint to the image restoration by analyzing the gradient distribution histogram of the images during the deblurring process. A variable substitution scheme was also established to simplify the compute. The third one is an improved RL Algorithm based on local prior, which can suppress the ringing artifacts caused by the failures in blur kernel estimation. To find the smooth region, we comput the standard deviation of pixels in a local window, and the image gradient in the region is constrained to make its distribution consistent with the deblurring image gradient.Finally, we present an algorithm for removing the motion blur from a single image by combining the two steps. To constrain the kernel, the algorithm chooses the spare regularization term and continuity regularization term of motion kernel. And the same time, to constraint the image it chooses the Gaussian noise distribution, the gradient distribution histogram of nature images and local prior of image.We demonstrate the effectiveness of our method from qualitative results. But this approach need be improved further to get better image result by solve the problem of parameters and the complex models.