基于熵正则L0梯度最小化模型的图像平滑方法

Image smoothing method based on entropy regular L0 gradient minimization model

图像平滑是计算机图像和视觉领域中的一项基本任务,L0梯度最小化模型是该领域效果较好的图像平滑处理方法之一。但是该方法存在着严重的阶梯效应,且缺乏对噪声的鲁棒性。为了克服这些缺点,本文提出了一种基于熵正则的L0梯度最小化模型的图像平滑方法。首先,采用快速局部均值滤波算法预处理图像,并将处理后的图像运用到L0梯度最小化模型中,以此减少噪声点对图像平滑的影响;然后,为更好地刻画处理后图像与原始图像的相似度,保护其边缘信息,引入熵因子作为模型正则项,以减轻阶梯效应对图像平滑效果的影响;最后,运用交替迭代寻优方法,求解能量函数的最优解,继而得到最终平滑图像。为验证所提方法的有效性,利用大量图像进行实验,实验结果表明:与L0梯度最小化模型、RTV模型、DTV模型、Superpixel L0模型相比,所提模型能够获得更好的平滑效果的同时,较好地克服阶梯效应,且对噪声的鲁棒性也有一定程度的提高。

Image smoothing is a basic task in the field of computer imaging and vision. The L0 gradient minimization model （LGM） is one of the better image smoothing methods in this field. However, the method has a serious staircase effect and lacks Robustness to noise. In order to overcome these shortcomings, the paper proposes an improved L0 gradient minimization model. Firstly, a fast local mean filter algorithm is used to preprocess the image, and the processed image is applied to the L0 gradient minimization model to reduce the influence of the noise point on the image smoothness; and then, to describe the similarity between the processed image and original image better, introducing the entropy factor as a model regular term, and reduce the effect of the staircase effect on the image smoothing effect; Finally, solve the energy function by alternative iterative optimization method, and then get the final smooth image. In order to verify the effectiveness of the proposed method, a large number of images were used for experiments. The experimental resuhs show that compared with the L0 gradient minimization model, RTV, DTV, Super, the proposed model can achieve smoothing effect and overcome the staircase effect better, one the other hand, to a certain extent,the robustness to noise has also been improvd.