基于BLMD和NSDFB算法的红外与可见光图像融合方法

针对传统图像融合方法容易导致融合图像出现细节不明显和目标信息不完整的问题, 本文提出一种基于二维局部均值分解 (Bidimensional Local Mean Decomposition, BLMD) 和非下采样方向滤波器组 (Nonsubsampled Directional Filter Banks, NSDFB) 算法的红外与可见光图像融合方法 (基于方向滤波的二维局部均值分解法, BidimensionalLocalMeanDecompositionbasedDirectionalFilteringAnalysis, BLMDDFA) .首先, 计算两幅原始图片的熵值, 同时提取熵值较大的图片的残余分量, 该残余分量与另一张原始图片有着较强的相关性.然后, 通过BLMD和NSDFB算法将残余分量和熵值较小的原始图片分解成低频子带和一系列不同尺度的高频方向子带, 并使用不同的融合规则分别对低频子带和高频子带进行融合.最后, 通过相应的逆变换运算获得融合图像.实验结果表明, 本文方法的融合性能在对比度、细节信息展示和目标突出方面均高于经典的融合算法, 在信息熵、标准差、空间频率和平均梯度方面较Laplacian方法中各指标分别提高了5.6％、28.9％、37.4％和47.6％, 信噪比较Laplacian方法降低了8.5％.

Infrared and Visible Image Fusion Based on BLMD and NSDFB

Because traditional image fusion methods can easily cause blurred details and dim targets,a new fusion approach based on bidimensional local mean decomposition(BLMD)and nonsubsampled directional filter banks(NSDFBs)for visible-infrared images is proposed.In this fusion framework,the entropies of two source images are first calculated,and the residue of the image whose entropy is larger is extracted,which is highly relevant for the other source images.Then,the residue and the other source image are decomposed into low-frequency subbands and a sequence of high-frequency directional subbands in different scales by using BLMD and NSDFBs.At the fusion stage,two relevant fusion rules are used in low-frequency subbands and high-frequency directional subbands,respectively.Finally,the fused image is obtained by applying the corresponding inverse transform.Experimental results show that the proposed fusion algorithm can obtain state-of-the-art performance for visible-infrared fusion images in the aspects of both objective assessment and subjective visual quality,even when the source images are captured in different conditions.Furthermore,the fused results have higher contrast,richer details,and more-remarkable targets than those of Laplacian image fusion methods,increasing by 5.6%,28.9%,37.4%,and 47.6%in the information entropy(IE),standard deviation(SD),spatial frequency(SF)and average gradient(AG),respectively,while decreasing by 8.5%in peak signal-to-noise ratio.