Motion-Blur Identification And Deblurring Using Single Image

A motion blurred image is produced as the scene and camera moved during the exposure time. It is a hot spot research on motion-blur images but often focused on image restoration. The motion blurred image implies more special physical information then the latent one, such objects'movement direction, speed, even acceleration.It is assured that the gradients in natural scenes obey heavy-tailed distribution. The different distribution characters between the latten image and the blurred image are analyzed in this paper. A useful method which is designed bases on the local windows'spectrum analysis is proposed to detect the partial motion-blur region. This method uses localized thinking, and considers the relationship between the spatial and frequency domain in the local window of pixels. Then we compute the ratio of high frequency coefficient and low frequency coefficients of each local window to determine the center pixel of the local window whether clear or blur. The experiments show that the detected regions can be used for objects motion tracking and deblurring, etc.The total motion-blur image has a spatially-invariant kernel. Many motion deblurring algorithms have been proposed in recently years. Such as heavy computation, non-robust and ringing are still limitations in most algorithms. In this paper, a blind deburring algorithm is depicted, which is effective and ringing-suppressed using single motion blurred image. The blind deburring algorithm contains two phases, blur kernel estimation and non-blind deconvolution as follow. Firstly, shock filter is used to predict the strong edges of latent image after noise suppression by the bilateral filter first. Then the predicted strong edges are used to estimate the blur kernel. The strong edges and the blur kernel will be reconstructed by using iteration steps.Secondly, the final estimated blur kernel is used to recovery the latent image by a non-blind deconvolution algorithm which is based on the hyper-Laplacian model. An alternating minimization scheme is adopted to optimize the energy equation, including a look-up table(LUT) is used for rapidly solving the image deconvolution. The fast Fourier transform is also considered to alternative the convolution operation to promote the speed.The experiments show that our blind deblurring algorithm can remove motion blur and improve image quality quickly and effectively by using single image. The non-blind deblurring algorithm yields effective result with medium size images in seconds. Both in the restoration effect and suppression of ringing, the method given in this paper yields the best overall performance, which is compared with the classical algorithms, such as Wiener and Richardson-Lucy algorithms.