摘要
对于同一场景可见光和红外图像提取的灰度特征存在差异导致匹配性能下降的问题,利用了该两个谱段图像仅在形状上模糊相似的特点,提出采用形状上下文思想实现红外与可见光图像之间的匹配方法。具体方法是在提取图像边缘的基础上,用采样模板获取点特征图像,然后利用提出的不等分坐标空间提取局部形状统计信息的直方图,再以形状上下文特征开支与全局特征匹配开支的加权和作为相似性测度实现匹配。与已有的形状上下文算法相比,此方法减少了25%的特征维数,鲁棒性更好。针对多个红外和可见光图像序列的匹配实验结果显示,此方法与归一化灰度互相关、Hausdorff距离、文献中的形状上下文和SIFT算子等匹配算法相比,其匹配正确率分别平均提高了34.7%、21.2%、7.7%、88.9%,证明了该方法的有效性。
To the problem that the matching performance between infrared images and visible images declines sharply due to the quite difference in gray features for pairs of them, an approach on shape contexts was adopted for matching between them based on the property that the object＇s shape in pairs of infrared images and visible images was rough similarity. Sampling template was applied to obtain images whose edge features were extracted, point-feature images were gained. And then, a new coordinate was proposed to calculate point-feature images＇ local shape information histograms. Finally, the weighted sum of the cost of shape contexts for matching points and the cost of total feature matching were used as the similarity measure for matching. Compared with the existed shape contexts algorithm, the feature dimensions by proposed method may reduce 25% and the robust is better. For large numbers of infrared and visible image sequences, the experimental results show that the average correct matching rate of this method increases 34.7% ,21.2% ,7.7% and 88.9% more than that of normal cross-correlation method, Hausdorff distance method, the existed shape contexts algorithm and SIFT algorithm, respectively, it demonstrates this method is effective.
出处
《红外与激光工程》
EI
CSCD
北大核心
2008年第6期1095-1100,共6页
Infrared and Laser Engineering
基金
总装预研重点基金项目(9140A01060507JW0506)
国防基础研究项目(A1420061266)
关键词
多谱图像
景象匹配
形状上下文
Multi-spectral images
Scene matching
Shape context
