摘要
提出一种改进的基于相似性约束的人脸超分辨率重建算法,采用迭代计算的方式将训练过程和学习过程整合在一起。首先从训练集中遴选出与待重建人脸最相似的训练库人脸参与迭代过程,随着迭代次数的增加,重建得到的高分辨率人脸越来越接近于原始高分辨率人脸;其中每次迭代分别统计待重建低分辨率人脸和训练集本次迭代参与的低分辨率人脸的相似性以及与训练集本次迭代参与的高分辨率人脸在局部结构上的相似性,以减少流形学习中低维空间到高维空间的一对多映射的限制。实验结果表明,与其他算法相比,文中所提的人脸重建算法不仅具有较低的空间复杂度,并且具有更好的主观和客观效果。
An improved face Super-Resolution （SR） reconstruction algorithm based on similarity constraints is proposed. The proposed algorithm incorporates training stage and learning stage together. Select the most similar face sets （ low resolution faces and corresponding high resolution faces） from the whole training face sets with the input Low Resolution （LR） face. With the increasing of iterative numbers ,the reconstruction result gets more and more close to the original High Resolution （HR） face. During each iterative leafing,the similarity between the input LR face image and the training LR face image is computed as well as the local structure similarity between the input LR face and the training HR face. The experimental results demonstrate that the proposed algorithm not only occupies less space complexity but also produces better subjective and objective results compared with other leading super-resolution reconstruction algorithms.
出处
《计算机技术与发展》
2015年第8期58-61,66,共5页
Computer Technology and Development
基金
国家自然科学基金资助项目(60802021
61172118
61271240)
江苏省高校自然科学重点研究项目(13KJA510004)
江苏省自然科学基金青年基金(BK20130867)
省属高校自然科学研究项目(12KJB510019)
关键词
迭代
相似性约束
流形学习
人脸重建
iteration
similarity constraints
manifold learning
face reconstruction
