基于原—对偶混合梯度下降法的图像恢复算法

Image restoration algorithm based on primal-dual hybrid gradient descent method

对基于原—对偶混合梯度下降法的图像恢复算法进行了改进。在原算法中,步长参数的设计对其恢复效果和收敛速度影响较大。为了改善算法性能,通过引入中间变量改变算法的形式,然后分离对偶向量的分量,用加权矩阵取代原算法中的步长参数,并对其分开设计。数值实验表明,在峰值信噪比(PSNR)和视觉效果相当的前提下,和原算法相比,降低了参数设计对算法性能的影响,且计算时间减少50%左右。

An improved algorithm for image restoration was proposed based on Primal-Dual Hybrid Gradient Descent （PDHGD） method. The preferences have a great impact on convergence rate of the known algorithm. The form was changed by introducing new variable, and the elements of the dual vector of the primal-dual hybrid model were separated, and then stepsize was replaced by using parameter matrices. The numerical experiments show that the improved algorithm has advantage on choosing parameters compared to the known algorithm, and the iterative number and CPU＇ time nearly declined by 50%, and at the same time, the improved algorithm has exactly the same effect on image restoration.