全局和局部结构内容自适应正则化的单幅图像超分辨模型

Global and local structural content adaptive regularization model for single image super-resolution

目的基于正则化的重建是单幅图像超分辨的重要方法之一。其中,如何构造合适的图像先验,增强超分辨重建过程中的边缘和纹理保持能力是该类方法的关键。提出一个全局和局部结构内容自适应正则化的单幅图像超分辨模型。方法该模型综合了图像梯度的全局非高斯性和局部结构方向自适应回归特性。首先,利用广义高斯分布拟合图像梯度模的重尾特性,由最大后验概率框架构造了图像全局内容感知的lα(0<α<1)范数稀疏性度量;然后,利用图像局部内容的各向异性相关性,给出基于Geman-Mc Clure(GM)权函数加权的局部结构方向自适应回归先验;最后利用半二次惩罚和变量分裂法,设计了该优化模型快速求解的超分辨算法。结果实验结果表明:在客观评价上,本文方法在峰值信噪比与结构相似度两方面优于现有的一些超分辨方法,在主观视觉效果上,能够很好的恢复图像的纹理细节和边缘信息。结论基于全局和局部结构内容自适应正则化的单幅图像超分辨方法在保持图像边缘和恢复图像纹理细节方面取得较好的重建性能。

Objective Regularization-based reconstruction is an important single-image super-resolution （SR） method. This class of methods aims to design effective image priors and incorporate them into a regularization framework to enhance edge- and texture-preserving capabilities during the SR process. Method In this study, a global and local structural content adaptive regularization model is proposed to solve the single-image SR problem. This model combines the global non- Gaussian statistics of an image gradient with the orientation-adaptive regression property of a local structure. Generalized Gaussian distribution is applied to fit the heavy-tailed distribution of the image gradient. A global content-based sparsity measure 1~ （ 0 〈 a 〈 1 ） norm is constructed under the maximum a posterior probability framework. The anisotropie corre- lation of local content is employed to construct an adaptive regression prior of the local structure based on the Geman-Mc- Clure function. Finally, a half-quadratic penalty method and a variable splitting technique are used to solve the model ef- fectively. Result For an objective assessment, experimental results demonstrate that the quality of the SR images obtained by the proposed method is better than those obtained by other methods in terms of peak signal-to-noise ratio and structural similarity. For a subjective evaluation, the proposed method can retain edge and image details effectively. Conclusion The proposed adaptive regularization method can preserve edge and image details effectively in single-image super resolution.