An Effective Total Variation Image Deblurring Method Based on Enhanced Sharp Edge Prediction

碩士 === 國立中央大學 === 資訊工程研究所 === 98 ===   Capturing satisfactory photographs by using a hand-held camera are quite challenging even under correct camera settings, in other words, unsatisfactory photographs will usually be taken under image blurs. The reasons that cause image blur are generally resulted...

Full description

Bibliographic Details
Main Authors: Tsung-yung Hung, 洪宗湧
Other Authors: Kuo-Chin Fan
Format: Others
Language:en_US
Published: 2010
Online Access:http://ndltd.ncl.edu.tw/handle/34349916721980758570
Description
Summary:碩士 === 國立中央大學 === 資訊工程研究所 === 98 ===   Capturing satisfactory photographs by using a hand-held camera are quite challenging even under correct camera settings, in other words, unsatisfactory photographs will usually be taken under image blurs. The reasons that cause image blur are generally resulted from camera shaking during exposure time which leads to inevitable important information loss and noise arisen.   Conventional blind deconvolution methods typically assume spacial/frequency domain constraints on images and/or parametric forms for the motion path of camera shaking for solving the ill-posed problem. The issues of current deblurring methods include reducing the common artifacts which are caused by image noise and/or errors in the estimated blur kernel while decreasing computation time. In this thesis, we present a novel blind deconvolution method which deblurs a single observed blurred image to produce a high-quality image. The purpose of our work is to suppress ringing artifacts effectively; moreover, to accelerate both blur kernel estimation and latent image restoration in iterative deblurring processes than conventional methods.   To reconstruct enhanced sharp image, we first apply some image pre-processing techniques and an edge mapping method to retain accurate edges and smooth regions from the estimated latent image, which will both join the later blur kernel estimation. Then, we formulate the optimization function by introducing penalty techniques with image derivatives to accelerate both the blur kernel estimation and latent image restoration. For latent image restoration, we also use auto-thresholding algorithm to truncate errors of the estimated kernel from the previous blur kernel estimation for recovering a sharp latent image while suppressing ringing artifacts.   Experimental results demonstrate that our proposed method is faster and more efficient than conventional methods for deblurring from a single blurred image. Moreover, high-quality images can be effectively and accurately produced with merely slight ringing effects.