摘要
提出基于稀疏表示和近邻嵌入的单帧图像超分辨率重构算法;为低分辨率和高分辨率图像块训练两个基于稀疏表示的过完备字典,在训练的低分辨率图像块和高分辨率图像块中分别选取与这两个字典原子最近的图像块近邻,通过图像块近邻来计算构图像块的权重;一旦得到权重矩阵,高分辨率重构图像块可以由低分辨率图像块与相应权重相乘来表示;与之前的算法相比,所提出的算法在计算字典原子与图像块距离的时候不是逐个图像块进行计算,而是先将图像块聚类,计算字典原子与类中心的距离,在距离最近的一类中选取图像块;计算权重矩阵的时间可以大大减少,提高计算效率;所得到的PSNR与其它算法相比,也有一定提高。
A single frame image super-resolution reconstruction algorithm based on sparse representation and neighbor embedding was proposed.Two complete dictionaries based on sparse representation were trained for low and high resolution image patches,in which the closest image patches to the two dictionary atoms were chosen.The weight of reconstructed image patches was represented by image patches neighbor.Once weight matrix was gotten.High resolution image patch can be expressed as low resolution image patch multipling by the corresponding weight.Compared with previous algorithms,when calculating the distance between the dictionary atoms and image patches,the proposed algorithm is not each image patch to calculate.Instead image patches are clustered,and calculate the distance between dictionary atoms and the clustering center,then select image patches in the closest category.Calculated time of weight matrix can be greatly reduced,and improve the computational efficiency.The resulting PSNR compared with other algorithms,there are also improved obviously.
出处
《计算机测量与控制》
2016年第5期173-177,共5页
Computer Measurement &Control
关键词
超分辨率重构
稀疏表示
过完备字典
图像块近邻
权重
super-resolution reconstruction
sparse representation
complete dictionaries
neighbor embedding
weight