| 引用本文: | 陈蔓,钟勇,李振东.基于SIFT字典学习的引导滤波多聚焦图像融合[J].哈尔滨工业大学学报,2018,50(11):59.DOI:10.11918/j.issn.0367-6234.201806014 |
| CHEN Man,ZHONG Yong,LI Zhendong.Multi-focus image fusion based on SIFT dictionary learning and guided filtering[J].Journal of Harbin Institute of Technology,2018,50(11):59.DOI:10.11918/j.issn.0367-6234.201806014 |
|
| 摘要: |
| 目前多数多聚焦图像融合算法仅仅针对解决某一类问题，如融合结果的局部细节保留能力差、空间连续性不足和对未配准的源图像鲁棒性差等.为能够同时解决以上问题，提出了一种基于SIFT字典学习的引导滤波多聚焦图像融合算法.该算法通过学习子字典克服了图像低秩表示具有全局性而局部细节描述不足的缺陷，同时子字典的分类利用图像SIFT特征的平移不变、尺度不变等特性，消除了未配准源图像融合结果出现伪影的现象.此外，在源图像的低秩表示系数融合过程中引入引导滤波，增加了融合图像的空间连续性.引导滤波的窗口大小是根据特征内容和非特征内容进行自适应选取，即属于特征内容的点选取较小的窗口，而属于非特征内容的点选取较大的窗口.为验证算法的有效性，实验过程中选取6组数据，包括3组广泛应用于研究的多聚焦图像以及3组实际拍摄的多聚焦图像.实验结果表明，该算法从主观视觉效果的定性分析和客观融合质量评价的定量分析都优于当前主流的多聚焦图像融合算法.
|
| 关键词: 图像融合 SIFT 低秩表示 字典学习 引导滤波 |
| DOI：10.11918/j.issn.0367-6234.201806014 |
| 分类号:TP399 |
| 文献标识码:A |
| 基金项目:四川省科技厅科技成果转化项(2014CC0043); 四川省科技创新苗子工程项目(SCMZ2006012) |
|
| Multi-focus image fusion based on SIFT dictionary learning and guided filtering |
|
CHEN Man1,2,ZHONG Yong1,LI Zhendong1,2
|
|
(1. Chengdu Institute of Computer Applications, Chinese Academy of Sciences, Chengdu, 610041; 2. University of Chinese Academy of Science, Beijing 100049)
|
| Abstract: |
| In order to solve the problem that the local detail retention ability, spatial continuity and non-registration problems of most multi-focus image fusion algorithms cannot be improved at the same time, this paper proposes a multi-focus image fusion algorithm based on SIFT dictionary learning and guided filtering. The algorithm overcomes the problem that the low rank representation of image can capture the global structure but could not preserve the local structure by learning sub-dictionaries. The classification of the sub-dictionaries utilizes the translation invariance and the scale invariance etc. of SIFT to eliminate the fusion artifacts of unregistered images. In addition, the adaptive-window guided filtering is performed during the low rank representation coefficients fusion progress, which increases the spatial continuity of fused image. Pixels with rich texture assign to small window, while weak texture pixels choose large window. We select 6 groups of data, including 3 groups of widely used images and 3 groups of real-world images for verifying the validity of the proposed algorithm. Experimental results show that this algorithm outperforms the current mainstream multi-focus image fusion algorithms from qualitative analysis and quantitative analysis.
|
| Key words: image fusion SIFT low rank representation dictionary learning guided filtering |
