摘要
提出了一种基于机器学习的超分辨率(SR)改进算法。首先建立一个包括低分辨率(LR)图像及其相应的高分辨率(HR)图像的训练样本集,为LR图像提供了HR的图像解释。把训练集中的每一幅图像分成若干个图像块,每一个图像块作为马尔可夫随机场(MRF)模型的结点,MRF模型参数从这些训练样本中学习得到,通过对训练样本中的LR图像块进行k-均值聚类减少计算开销,并用k-均值的聚类结果提出了一种新的相容函数形式。实验结果表明,该算法是可行的,并与同类算法相比能取得较好的结果,使得SR后的图像更平滑自然。
An improved algorithm is presented for super resolution based on machine learning.A training sample set is set up which contains low-resolution images and the corresponding high-resolution images.These training samples provide high-resolution image interpretation for the low-resolution images.Every image in the training set is divided into several patches and each of them is assigned one node of a Markov Random Field（MRF）.The parameters of MRF are learned from these training samples and the probability distribution is computed by k-means cluster algorithm,which can greatly reduce computing cost.And a novel form of compatibility function is proposed using the result of k-means algorithm.The result of experiment demonstrates that our improved algorithm is feasible and lays over other algorithms.It also makes the super-resolved image more natural and smooth.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2010年第1期120-123,共4页
Journal of Optoelectronics·Laser
基金
国家自然科学基金资助项目(60872064)
天津市自然科学基金资助项目(08JCYBJC12200
08JCYBJC12300)
