In the field of variational image processing, more and more experiments have confirmed the effectiveness of nonconvex models for characterizing practical problems. Therefore, it is urgent to design the fast optimization method with convergence guarantee for solving the nonconvex variational models. Although there have been some theoretical results in the nonconvex optimization method, the assumptions are strict and further research is needed. This project aims to study the nonconvex variational model and primal-dual algorithm in image restoration problems, including three typical nonconvex variational models and difference of convex functions algorithm, alternating direction method of multipliers, primal-dual hybrid gradient method. The convergence speed is improved by using the techniques of proximal points, continuity strategy, adaptive parameters and the convergence of the new method under weaker conditions is proved by the convex analysis, dual theory and Kurdyka-Lojasiewicz properties. Noise remove problems and medical reconstruction problems with a small amount of data are tested. Practical and efficient mathematical software will be developed. This project is scientific significant and extremely valuable, not only for providing new methods for image restoration problems, but also for promoting the development of nonconvex optimization theory.
在变分图像处理领域中,越来越多的实验证实了非凸模型对于刻画实际问题的有效性。因此,如何设计快速收敛的优化方法求解这些非凸变分模型成为亟待解决的问题。尽管前期对非凸优化方法已经有一些理论结果,但假设条件要求严格,还需进一步研究。本项目拟开展图像恢复问题中非凸变分模型及其原始-对偶算法的研究,主要包括三类典型的非凸变分模型和对应的凸函数差方法、交替方向乘子法和原始-对偶混合梯度法。利用邻近点、连续化策略与自适应参数技巧,提高收敛速度。借助凸分析、对偶理论及Kurdyka-Lojasiewicz性质,减弱算法收敛的条件。对新方法在噪音去除和少量数据医学图像重建问题上进行数值实验,编写实用高效的数值程序包。该项目的实施不仅能为图像恢复问题提供新方法,也可促进非凸优化理论的发展,具有重要的科学意义和实用价值。
本项目对图像恢复问题中的非凸变分模型及其原始-对偶类方法尝试一些新研究,建立原始-对偶类非凸优化方法收敛性相关理论,有效解决高斯、脉冲和泊松噪音下的图像恢复或重建问题。项目主要研究了三类典型的非凸变分模型和交替方向乘子法、凸函数差方法及原始-对偶混合梯度法。取得成果包括:a) 针对凸数据拟合项加非凸变分正则项图像恢复问题,研究了求解此非凸模型的交替方向乘子法,并在较弱条件下证明了算法的收敛性,算法应用到不同图像恢复问题中,包括高斯噪音去除,去模糊、MRI、CT医学成像和图像超分辨等;b) 针对非凸数据拟合项加凸或非凸变分正则项的脉冲噪音图像恢复问题,提出了可以转化为凸函数相减形式的非凸变分模型,如数据拟合项为SCAD函数、正则项为全变分或对数全变分,研究了求解此类模型的邻近凸函数差方法,并分析收敛性;c) 针对泊松噪音去除问题,利用凸变分正则项的对偶函数,转化原始问题为极大-极小鞍点问题,拓展凸问题的原始-对偶混合梯度方法求解此模型。 项目立足国际前沿的创新研究,为图像恢复提供有效方法和理论保证。