基于L₀正则化的文本图像去模糊方法研究

【作者】 张红；

【导师】 刘辉；

【作者基本信息】 昆明理工大学 ， 计算机技术（专业学位）， 2017， 硕士

【摘要】 文本图像去模糊包括纯文本图像去模糊和非纯文本图像去模糊,然而图像的去模糊过程是一个不适定性问题,不能直接得到有效解,所以采用Lp范数正则化的方法来求解,具体内容:分析TV正则化模型可知,该模型中保真项比较单一,正则项比较丰富。而正则项中L0范数是一个NP组合优化问题,很难对其进行求解。虽然往往可以将L0范数问题转换成L1范数进行近似求解,但L1范数通常情况下并不能得到最稀疏的解。其次,将已存在的图像去模糊方法应用于文本模糊图像中,也不能取得良好的效果。针对上述问题,为了提高文本图像去模糊的质量,设计了基于L0正则化的文本图像去模糊方法。对于数据保真项,采用了基于文本像素的强度特征和梯度特征作为先验知识的方法,该方法能够有效的利用L0范数正则化来约束文本图像的强度特征,且使用文本图像的梯度特征来保持文本图像的边缘变化情况。对于L0最小正则化问题,使用了最新高效的半二次分裂方法进行求解,并在半二次分裂中加入了中间变量与交替迭代方法,使得去模糊效果更佳,同时,使用中间变量的像素梯度方法进行模糊核的估计。实验证明,与现有的文本图像去模糊方法相比,本文中的文本去模糊方法能取得更好的效果,同时,能够保持文本图像的边缘细节,从而在主观和客观上验证了本文方法的优异性。更多还原

【Abstract】 The Scene text image deblurring includes texture deblurring and non-texture deblurring.However,the image deblurring is a ill-posed inverse problem.Solving this problem directly,cannot get accurate solution.So,the Lp regularization method will be used to deal with this in the text image deblurring.The main works as following:Analysis of TV regularization model shows that the fidelity items is relatively simple,and the regularization items are relatively rich.The Lo norm in the regularization term is an NP combination optimization problem,which is difficult to solve.Although the L0 norm problem can often be solved by the approximate L1 norm,but the L1 norm is usually not the most sparse solution.Secondly,the existing image deblurring method,which applied to text blurred images,and can not achieve good results.Aiming at the above problem,in order to improve the quality of text image deblurring,designed a text image deblurring method based on Lo regularization.For the data fidelity term,the intensity feature and gradient feature based on the text pixel are used as a prior knowledge,which method can effectively use the Lo norm regularization to constrain the intensity characteristics of the text image,and can use the gradient feature of the text image to keep the edge changes of the text image.For the Lo minimum regularization problem,the most efficient half-quadratic splitting method is used to solve this problem,and the auxiliary variables and the alternating iterative method are added to the half-quadratic splitting method,which makes the deblurring results better,at the same time,using the pixel gradient of the intermediate variable to estimate the image blur kernel.Compared with the existing text image deblurring method,the experiment proves that the text deblurring method in this paper can achieve better deblurring effect,and can keep the edge detail of the text image,thus verifying the excellence of article method from subject and object.更多还原