摘要：

针对传统基于超完备字典的图像超分辨率重建算法训练样本庞大、训练时间长、稀疏度固定,且迭代时间长的问题,提出一种快速的图像超分辨率重建算法。该算法在字典训练阶段引入快速核密度估计算法对训练样本规模进行估计,得到数量合理的训练样本,在稀疏表示阶段使用改进的广义正交匹配追踪算法,克服稀疏表示算法中固定稀疏度的缺陷。实验结果表明,相比传统字典训练算法,该算法能提高超分辨率重构的精度,且平均迭代时间较少。

关键词: 稀疏表示, 压缩感知, 快速核密度估计, 广义正交匹配追踪, 超分辨率, 字典学习

Abstract:

The traditional Super Resolution ( SR ) algorithm via over-complete sparse representation has several problems,such as too large training patches,long training and iteration time,and fixed sparse degree. In view of these disadvantages,a fast SR algorithm is proposed. The core of this algorithm is to estimate the scale of the training patches by introducing Fast Kernel Density Estimation(FastKDE) to get the reasonable number of training patches in the stage of dictionary learning,and to overcome the shortcomings of greed series of sparse representation algorithms with fixed sparse degree and shortens the iteration time by using improved Generalized Orthogonal Matching Pursuit(GOMP) algorithm in the stage of sparse representation. Experimental results show that compared with the traditional dictionary training algorithm,this algorithm can improve the accuracy of SR reconstruction,and the average iteration time is less.

Key words: sparse representation, compressed sensing, Fast Kernel Density Estimation ( FastKDE ), Generalized Orthogonal Matching Pursuit(GOMP), Super Resolution(SR), dictionary learning

中图分类号: 

• TN911. 72