基于异方差TR约束的自适应人脸超分辨率

Adaptive face super resolution based on heteroscedasticity Tikhonov regularization constraints

为解决基于局部图像块的人脸超分辨率表达系数实际的异方差特性,提出一种基于异方差TR(Tikhonov regularization)约束的自适应人脸超分辨率方法。同时对重构误差和表达系数进行加权约束,将同方差推广到异方差,获得更加准确的自适应先验;在此基础上,通过上下文信息块丰富基于位置块的先验,使用残差学习预测更精准的高频信息,获得更好的重建性能。在CAS-PEAL-R1人脸数据库上的实验结果表明,该方法优于其它对比方法,包括基于卷积神经网络的超分辨率算法。

To solve the actual heteroscedasticity of face super-resolution representation coefficients based on local image patches,an adaptive face super-resolution method based on the heteroskedastic TR(Tikhonov regularization)constraint was proposed.Both the reconstruction error and the representation coefficient were weighted and constrained,and the homoscedastic was extended to the heteroscedasticity to obtain a more accurate adaptive prior.On this basis,the location patch-based priori was also enriched through context information patches,and residual learning was used to predict more accurate high-frequency information for better reconstruction performance.Experimental results on CAS-PEAL-R1 face database show that the proposed method is superior to other methods,including super-resolution algorithm based on convolutional neural network.