摘要
研究了软硬件相结合的亚像元超分辨成像技术。首先通过探测器扫描获得同一场景彼此错位亚像元像素的多帧图像作为训练样本和输入图像;然后针对传统的局部线性嵌入(LLE)实例学习超分辨算法过于依赖外部训练样本,不利于光电成像系统直接处理等缺点,提出了一种基于自学习的改进LLE算法;采用新的LLE权值计算方法获得正数权值,同时对初始估计再次运用自学习LLE方法恢复丢失的高频细节信息。仿真实验结果表明,该算法重构的图像的信噪比比传统LLE超分辨算法提高了0.8dB,运行时间提高了75%,视觉上可感知重构图像的细节信息更丰富。与其它方法相比,用搭载的微位移实验平台运行本文算法所获得重构图像的信噪比和信息熵都有很大提高,表明本文算法能获得高质量和高分辨率的重构图像。
A super-resolution technology of combining hardware and software was researched. Firstly, the detector scanning was used to obtain multiple images with the same scene produced by different motion parameters and they were chosen to be training sets and input images. In consideration of that traditional Local Linear Embedding （LLE） super-resolution technology is over-relying on external training images and is inconvenient for processing image directly, a improved self-learning algorithm based on the LLE was proposed. The new LLE weight calculation method was proposed to obtain initial estimation of HR image. Meanwhile, self-learning LLE algorithm was used to recover lost high-frequency information of initial estimation and to obtain the final estimation. Simulation results show that the Peak Signal to Noise Radio（PSNR） of the reconstructed image by proposed algorithm improves 0.8 dB and operation time shortens by 75% as compared with those of conventional LLE method, respectively. Moreover, in the real scene experiment of micro-displacement platform, the Signal to Noise Radio （SNR） and information entropy of the reconstructed image by proposed algorithm have also greatly improved as compared with those of other algorithms. The algorithm provides high quality reconstruction image and improves the resolution of the captured image.
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2015年第9期2677-2686,共10页
Optics and Precision Engineering
基金
吉林省重大科技攻关项目(No.11ZDGG001)
国家自然科学基金青年基金资助项目(No.60902067)
关键词
超分辨成像
亚像元图像
自学习
局部线性嵌入
训练样本
super-resolution imaging
sub-pixel image
self-learning
Local Linear Embedding（LLE）
training set
