摘要
针对传统基于学习的人脸图像超分辨率算法存在高频细节信息损失过多问题,提出一种全局重建和位置块残差补偿相结合的人脸图像超分辨率新算法.首先利用高、低分辨率训练集所有样本,使用基于权值学习的全局重建算法得到初步的人脸图像,再结合图像模糊和下采样过程,产生高、低分辨率残差图像训练集,最后使用基于位置块的残差补偿算法,对初步的人脸图像进行高频细节补偿得到最终结果.对比实验结果表明,相比同类基于学习的人脸图像超分辨率算法,在将人脸图像分辨率提高4×4倍的情况下,新算法的平均峰值信噪比可提高0.65～3.55dB,可以更好地重建出局部高频细节信息.
Focusing on the high loss of subtle facial details in traditional algorithm, a face image super-resolution algorithm through global reconstruction and position-patch based residue compensation is presented. The optimal coefficients of the low-resolution training images are computed and transformed into high-resolution space to reconstruct the global high-resolution image. The residue training set is obtained by a smoothing and down-sampled processing. The residue compensation based on position is performed to better recover face subtle details using the residue training set. Experimental results show that the proposed approach synthesizes high-resolution faces with more details and the average of peak signal-to-noise ratios is improved about 0. 65 dB to 3. 55 dB compared with some existing learning-based methods.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2010年第4期9-12,共4页
Journal of Xi'an Jiaotong University
基金
国家高技术研究发展计划资助项目(2007AA01Z176)
国家自然科学基金资助项目(60972124)
关键词
人脸图像
超分辨率
残差补偿
位置块
face image
super-resolution
residue compensatiom Dosition-oatch
