深度协作表达的人脸超分辨重建算法  

Face Super-Resolution by Deep Collaborative Representation

作　　者：卢涛[1,2] 潘兰兰 管英杰[1,2] 曾康利[1,2] Lu Tao;Pan Lanlan;Guan Yingjie;Zeng Kangli(School of Computer Science and Engineering, Wuhan Institute of Technology, Wuhan 430205;Hubei Province Key Laboratory of Intelligent Robot, Wuhan Institute of Technology, Wuhan 430205)

机构地区：[1]武汉工程大学计算机科学与工程学院,武汉430205 [2]武汉工程大学智能机器人湖北省重点实验室,武汉430205

出　　处：《计算机辅助设计与图形学学报》2019年第4期596-601,共6页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61502354;61501413;61671332;41501505);湖北省自然科学基金(2015CFB451;2014CFA130;2012FFA099;2012FFA134;2013CF125);中央引导地方科技发展专项(2018ZYYD059);湖北省高等学校省级教学研究项目(2017324);武汉工程大学重点教学建设工程项目(Z2017009)

摘　　要：为了获得更高质量的高分辨率人脸图像,提出一种基于深度协作表达的人脸超分辨率算法.该算法首先对训练样本重叠取块和多方向梯度特征提取得到初始化字典;再对初始化字典进行逐层迭代更新,同时利用协作表达更新对应每层的最优表达权重系数;最后将最后一个表达层所有的重建块合成为最终高分辨率图像.在FEI和CMU FrontalFace数据集上的实验结果表明,与传统超分辨率重建算法相比,该算法提升了单层表达算法的精度,在主观和客观评价性能上均超过现有算法,甚至包括基于深度学习算法.In order to obtain high-resolution face images with better quality, a face super-resolution algorithm based on deep collaborative representation is proposed. Firstly, the algorithm extracts overlapping patches and multi-directional gradient feature to obtain the initialization dictionary. Then, the initialization dictionary is updated iteratively layer by layer, and the optimal expression weight coefficients corresponding to each layer are updated by the collaborative representation. Finally, all the reconstructed patches of the last layer are combined into a final high-resolution image. Compared with traditional super-resolution reconstruction algorithms, the experimental results on FEI and CMU Frontal Face datasets show that the proposed algorithm improves the accuracy of the single-level representation algorithm and outperforms the existing algorithms in both subjective and objective evaluation performance, even including the deep-based learning algorithm.

关 键 词：人脸图像 超分辨率 字典对学习 深度协作表达 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]