面向单幅图像去雨的非相干字典学习及其稀疏表示研究

发布时间：2018-09-07 10:50

【摘要】：提出一种非相干字典学习及稀疏表示方法,并将其应用于单幅图像去雨。该方法在字典学习阶段,为降低有雨原子与无雨原子间的相似性,引入字典的非相干性,构建新的目标函数,不仅可以保证有雨字典与无雨字典的可分性,而且学习的非相干字典具有类似于紧框架的性质,可以逼近等角紧框架。通过有雨字典与无雨字典对高频图像的稀疏表示,能够更好地分离出高频图像中的有雨分量与无雨分量,将高频无雨分量与低频图像融合实现图像去雨。采用合成雨图与真实雨图对算法进行验证,实验结果表明,算法所学习的非相干字典具有较好的稀疏表示性能,去雨后的图像雨线残留较少,边缘细节保持较好,视觉效果更为清晰自然。
[Abstract]:An incoherent dictionary learning and sparse representation method is proposed and applied to single image to remove rain. In order to reduce the similarity between rain-free and rain-free atoms, this method can not only guarantee the separability of Rain dictionaries and Rain dictionaries, but also introduce the non-coherence of dictionaries and construct a new objective function. Moreover, the learning incoherent dictionaries have properties similar to compact frames, which can be approximated to isobaric compact frames. The sparse representation of high frequency images by rain dictionary and rain free dictionary can better separate the rain component from the rain free component in the high frequency image, and combine the high frequency rain free component with the low frequency image to remove the rain. The synthetic rain map and the real rain map are used to verify the algorithm. The experimental results show that the incoherent dictionary studied by the algorithm has better sparse representation performance, the rain-line of the image after rain removal is less, and the edge details remain better. Visual effects are more clear and natural.
【作者单位】：湘潭大学信息工程学院;湖南大学电气与信息工程学院;湘潭大学控制工程研究所;
【基金】：国家自然科学基金资助项目(No.61573299,No.61673162,No.61672216,No.61602397) 湖南省自然科学基金资助项目(No.2017JJ3315,No.2017JJ2251,No.2016JJ3125)~~
【分类号】：TP391.41

【参考文献】