基于局部像素嵌入的人脸图像超分辨率重构  

Reconstruction of Super-Resolution Facial Image Based on Local Pixel Embedding

作　　者：胡宇[1] 沈庭芝[1] 刘朋樟[1] 赵三元[1] 

机构地区：[1]北京理工大学信息与电子学院,北京100081

出　　处：《北京理工大学学报》2011年第2期201-205,共5页Transactions of Beijing Institute of Technology

基　　金：国家自然科学基金资助项目(60772066)

摘　　要：利用像素点在邻域空间的线性嵌入关系作为先验约束来重构高分辨率(HR)人脸图像.算法从HR训练样本集中选择与输入人脸最相近的K个样本进行配准,并以配准后的样本作为参考,学习目标图像中像素点的局部嵌入系数.在学习过程中,算法通过自适应调整各参考样本的权重来减小配准误差的影响,并利用总变差最小化约束嵌入系数的平滑度.结合局部像素嵌入关系以及降质模型,算法可以在最大后验估计的框架下实现对目标人脸的超分辨率重构.实验表明,重建的HR图像拥有更加细腻、清晰的局部特征,其平均峰值信噪比和结构相似度分别比对比算法高出1.26dB和0.04.The local linear embedding relationship of pixels was used as prior to constrain the reconstruction of HR facial image.The proposed algorithm selected K neighboring face samples,which were more similar to the input face,from the training set,and then took them as reference examples after their registrations so as to learn the embedding coefficients of pixels in target HR image.The weights of respective face examples in the learning process were adaptively adjusted in order to reduce the influence of registration errors,and the total variation minimization was used to constrain the smoothness of embedding coefficients.Combining the local pixel embedding relationship and the degradation model,the method could reconstruct target HR image using the maximum posterior estimation framework.Experimental result shows that the reconstructed HR images can posses more delicate and clear local features,and the PSNR and SSIM are 1.26 dB and 0.04 higher than that of using comparison algorithms,respectively.

关 键 词：超分辨率重构 局部像素嵌入 总变差最小化 人脸图像 

分 类 号：TN911.73[电子电信—通信与信息系统]