基于单图像超分辨率的约束随机森林算法

Constrained random forest algorithm for single image super-resolution

为解决现有单图像超分辨率算法对不同类型图像鲁棒性不强的问题,提出一种基于多模糊核约束的随机森林算法。结合多模糊核扩展随机森林训练阶段的输入特征向量,由低分辨率图像块和对应的模糊核组合生成输入特征变量;将多模糊核引入到决策树构建的质量测度函数中,用于约束决策树构建时的结点划分,使生成的叶结点更纯;采用多模糊核对叶结点的回归模型进行约束,降低决策树的预测误差。仿真结果表明,与主流的基于学习的单图像超分辨率算法相比,该算法对不同图像的鲁棒性更强,采用该方法重建的超分辨率图像的峰值信噪比更高。

For solving the problem that the single image super-resolution algorithm is not robust to different images, a random forest algorithm based on multi-blur kernels constrained was proposed. The input feature vectors for training process of random forest were extended by combining multi-blur kernels, and input feature vectors combined by low resolution image patches and corresponded blur kernels were generated. The multi-blur kernels were introduced into the quality measurement function for building decision tree, to constrain the split nodes while building decision tree and to make the generated leaf nodes more pure. Multi-blur kernels were used to constrain the regression model of leaf nodes to reduce prediction error of the decision tree. Ex perimental results show that compared with the existing learning-based single image super-resolution algorithms, the proposed method is more robust to different images, and the peak signal to noise ratio of the generated super-resolution images using the proposed method is higher.