基于非下采样轮廓波变换与引导滤波器的红外及可见光图像融合

为克服传统融合方法对灰度相关性较弱的红外与可见光图像融合存在的不足，提出一种基于 非下采样轮廓波变换（NSCT）与引导滤波器的融合方法。利用NSCT对源图像进行多尺度多方向分解，分离出包含在不同频带内的特征信息，得到一个低频近似图像和多个高频方向细节图像；局域窗口加权平均能量和改进拉普拉斯能量分别作为低频近似图像的活性测度，构造显著特征图对近似图像进行加权平均，以解决能量保持和细节提取两个关键问题；在方向细节图像中，基于活性测度取大规则获得决策映射图，将源图像作为引导图、决策映射图作为输入图像进行引导滤波，得到权重分配图，对方向细节图像进行加权平均，降低噪声的敏感度。对融合后的近似图像和方向细节图像进行NSCT逆变换，得到最后的融合图像。采用多组红外与可见光图像进行融合实验，并对融合结果进行客观评价。实验结果表明，该融合方法在主观和客观评价上均优于已有文献的一些典型融合方法，如基于两尺度分解的引导滤波融合方法、NSCT域内稀疏表示融合方法、基于像素显著性的交叉双边滤波融合方法、基于深度学习的卷积神经网络融合方法、基于显著性检测的双尺度融合方法，可获得更好的融合效果。

Infrared and Visible Images Fusion based on Non-subsampled Contourlet Transform and Guided Filter

An image fusion method based on non-subsampled contourlet transform (NSCT) and guided filter is proposed to overcome the shortcomings of traditional fusion methods for infrared and visible images with weak gray correlation. The source images are decomposed into multi-scale and multi-directional sub-bands using NSCT, which can separate the feature information contained in different frequency domains to obtain a low-frequency approximate image and the high-frequency directional sub-band images. The local window weighted average energy and sum-modified-Laplacian energy are regarded as the activity measures of low frequency approximate image, respectively, which are used to construct a salient feature map to solve the two key problems of energy preservation and detail extraction. In the directional sub-band image, the maximun activity measure rule is used to obtain the decision maps, the source images are taken as the guided images, and the decision maps are used as the input images for guided filtering. The weight distribution graph is obtained to weight and average the directional sub-band images to reduce the noise sensitivity. Finally, the fused approximate and directional sub-band images are reconstructed by non-subsampled contourlet inverse transform, and the final fusion image is obtained. Some fusion experiments on several sets of infrared and visible images were did, and the objective performance assessments were implemented to fusion results. The experimental results indicate that the proposed method performs better in subjective and objective assessments than a few existing typical fusion techniques，such as guided filter fusion method based on two-scale decomposition, sparse representation fusion method in NSCT domain, cross bilateral filter fusion method based on pixel saliency, convolution neural network fusion method based on deep learning, two-scale fusion method based on saliency detection, and obtains better fusion performance.