Contourlet变换应用于图像复原时容易引入伪吉布斯现象。非下采样Contourlet变换(NSCT)具有平移不变性, 能够克服伪吉布斯现象, 但是由于基于学习的超分辨率复原需要建立不同分辨率的关系, 而NSCT变换的结果是每一层图像大小都一样, 不能像拉普拉斯金字塔那样建立高低分辨率图像的对应关系及运算量较大。针对这些问题, 提出了基于改进的非下采样Contourlet变换(INSCT)的超分辨率复原算法。为了表示人脸特征, 算法首先建立了INSCT金字塔。然后针对人脸的特殊性, 在匹配过程中, 采用对应点进行匹配的方法。实验表明该算法具有较好的性能, 复原出的超分辨率人脸图像无论在主观视觉效果上还是在客观评价指标上都取得较好的结果, 复原的图像具有更好的视觉效果, 更逼真, 更接近于原始高分辨率图像。

Contourlet transform in image restoration is apt to bring pseudo-Gibbs phenomenon. Nonsubsampled Contourlet (NSCT) has better frequency selectivity and regularity compared with the contourlet transform, and it can overcome pseudo-Gibbs phenomenon. However, learning-based super-resolution need establish the relation of different resolutions. Different from Laplacian pyramid, result of NSCT at each layer is with the same size, but cannot establish multi-resolution pyramid, and it needs large amount of computation. According to the problem of NSCT in learning-based super-resolution, a face image super-resolution restoration algorithm based on INSCT is proposed. In order to represent face images features, the INSCT pyramid is established. For particularity of the face images, corresponding points in matching process are used. Experimental results show that this method has good performance, and achieves better results on subjective visual effects and on objective peak signal-to-noise ratio. And the results are closer to the real images.