复合正则化的图像超分辨率复原方法

Imagesuper-resolution via hybrid regularization methods

摘要:

图像超分辨率复原是计算机视觉领域的一个基础研究方向.为提高具有高度结构化特征图像超分辨率复原效果, 提出了一种基于复合正则化的超分辨率复原方法.该方法整合了多个不同种类范数的正则化项, 将具有旋转不变性的方向全变分组稀疏正则化(TI-DTV)方法以及小波分析方法嵌入至目标函数中.全变分组稀疏正则化(TI-DTV)方法是一种可以有效解决高度结构化图像中直线边缘区域超分辨率复原的方法, 但TI-DTV中的全变分(TV)和方向全变分(DTV)模型可能会导致图像阶梯化效应(staircase artifacts), 而小波分析项则可以提高图像纹理信息的复原效果, 可减小阶梯化效应的影响.为了解决不同范数下的混合正则化问题, 利用一阶对偶圆锥形解法(TFOCS)的思想, 推导出了一阶对偶形式的快速解法.结果表明, 在真实图像集的实验中, 通过与全变分、小波分析、TI-DTV等超分辨率复原方法的比较, 可以明显的看出该方法结果较其他方法更清晰, 对直线型结构复原效果有一定的提高, 同时保留了更多的细节信息, 峰值信噪比(PSNR)和结构相似性(SSIM)也有明显提高.

Abstract:

Transform-invariant group-sparse regularization with directional total variation is an efficient method to solve the super-resolution problem in the regions with highly structured straight edges which is deformed from the real urban scenes and space scenes. Total variation, however, probably leads to staircase artifacts, which may affect the result of super-resolution to large extent. In this paper, we present a new method by mixing different regularizers, especially by combining wavelet analysis with the Transform invariant directional total variation objective function. This allows us to simultaneously recover textures and local geometry structures, particularly highly structured straight edges. To solve this hybrid regularization problem with different norms, we used the idea of templates for first-order conic solvers, and derived the solution of the whole object function that we proposed. Experiments on same real image collections show that our method is more effective than prior works.