基于在线字典学习的人脸超分辨率重建  

Human face super-resolution reconstruction based on online dictionary learning

作　　者：刘芳华[1] 阮若林[1] 王建峰[1] 倪浩[1] 

机构地区：[1]湖北科技学院,湖北咸宁437100

出　　处：《现代电子技术》2017年第13期57-61,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目(61271256);湖北省自然科学基金项目(2015CFB452);湖北省高等学校优秀中青年科技创新团队计划项目(T201513);湖北省教育厅科研计划指导性项目(B2015080)

摘　　要：针对基于学习的人脸超分辨率算法噪点、伪影较多,且噪声鲁棒性较差的问题,提出一种基于在线字典学习的人脸超分辨率重建算法。以人脸图集作为训练图库,运用在线字典学习方法提高字典训练的精度。独立调整字典学习阶段的正则化参数λt和求解重建稀疏系数阶段的λr,以获取最优的超完备字典和稀疏系数用于图像重建。实验结果表明,目标图像峰值信噪比比同一类型的稀疏编码超分法平均提高了0.85 d B,结构相似性增加了0.013 3,有效地抑制了噪点和伪影。在含噪人脸图像应用中,噪声水平提高时,峰值信噪比下降相对较平缓,提升人脸超分效果的同时改善了算法的噪声鲁棒性。Aiming at the problems of more noisy points and artifacts,and poor noise robustness existing in the learning-based human face super-resolution algorithm,a human face super-resolution reconstruction algorithm based on online dictionary learning is proposed. The human face image set is taken as the training library. The online dictionary learning method is used to improve the accuracy of dictionary training. The regularization parameter λt of the dictionary learning phase is regulated independently,and regularization parameter λr in the reconstruction stage of the sparse coefficients is solved to get the optimal overcomplete dictionary and sparse coefficients for image reconstruction. The experimental results show that the peak signal-to-noise ratio（PSNR）of the target image of the proposed algorithm is 0.85 d B higher and the structural similarity is 0.0133 higher than that of the same type sparse coding super-resolution algorithm averagely,which can restrain the noisy point and artifact effectively.The application result of noisy human face image shows that the PSNR is decreased smoothly when the noise level is increased,which can improve the robustness against noise while promoting the performance of face super-resolution.

关 键 词：在线字典学习 超分辨率重建 含噪人脸图像 稀疏编码 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]