Research On The Editing Techniques Of A Blurred Image

The relative motion or defocus in the process of image capture will cause image blur. Blur image restoration became one of the hotspots in computer vision with the widespread usage of digital camera. This paper researches several editing techniques of a single blurred image. Our main work is the following:First, We study partial blur and propose a novel method for detect and extract the blur region automatically. Inspired by the lazy snapping technique in the clear image matting research, the method mainly consists of two steps:rough positioning and fine extraction. For rough positioning, an algorithm for blur/non-blur region detection using image patches is proposed. It roughly recognizes the blurry regions through combining the Gaussian Mixture Model of image gradients in spatial domain and the statistical analysis of power spectrum in frequency domain. For fine extraction, an improved lazy snapping based on the result of rough positioning is presented for automatically and precisely extracting the blur area. Experimental results demonstrate that the efficiency of the proposed method.Second, we proposed an effective method to compute the direction motion blur. We present a detailed procedure using FFT based on the frequency properties of linear motion blur. The procedure includes an accurate angle computation method combining Radon transform and least squares fitting, and a blur length computation method using projective ID collapse of the specturm. In the procedure, a windowing step is adopted to eliminate the boundary artifacts, and the blur direction is computed via the line corresponding to the direction in the cepstrum. Experiments with known kernels demonstrate that our proposed method is effective.Third, we summarize the existing popular deconvolution algorithms. First, we discuss on several classic deconvolution methdos, and see that there are three relatively serious defects:ringing effect, noise amplification and overly smooth boundary. We further discuss several effective methods to solve these problems: residual RL deconvolution, Gaussian-constrined deconvolution, standard Laplacian and Hyper-Laplacian deconvlution model. Signal-to-noise ratio and recovery spend time are used as evaluation standards for their recovery results.