基于流形学习和梯度约束的图像超分辨率重建

将改进的基于流形学习的超分辨率重建与基于梯度约束的正则化重建结合起来,提出一种新的单帧图像超分辨率重建算法.该算法首先针对基于流形学习的超分辨率重建,提出新的特征提取方法,联合归一化亮度与平稳小波变换细节子带系数两个特征矢量,提高重建性能;然后将学习得到的高分辨率图像作为初始估计,将其梯度作为目标梯度域,进行基于梯度约束的正则化重建,得到最终的高分辨率图像.与现有的一些算法相比,文中算法无论在视觉效果还是客观评价上都具有较好的重建性能.

Super-Resolution Image Reconstruction Based on Manifold Learning and Gradient Constraint

Proposed in this paper is a novel super-resolution reconstruction algorithm of single-frame images,which integrates the improved super-resolution reconstruction based on manifold learning with the regularized reconstruction based on gradient constraint.In this algorithm,a new feature extraction method,which combines the two feature vectors of the normalized luminance and the detail sub-band coefficient of stationary wavelet transform, is put forward for the super-resolution reconstruction based on manifold learning,and is used to improve the reconstruction performance.Then,a regularized reconstruction based on gradient constraint is implemented to obtain the final high-resolution image,with the learned high-resolution image and its gradient respectively as the initial estimate and the target gradient field.As compared with some existing algorithms,the proposed algorithm is of better reconstruction performance in terms of both visual effect and objective evaluation.