基于显著性检测与梯度导向滤波的红外与可见光图像融合

Infrared and visible image fusion based on saliency detection and gradient domain guided image filtering

针对红外与可见光融合图像背景信息丰富度不足,以及在融合过程中红外目标显著性、边缘轮廓清晰度和细节纹理信息的保留难以同时兼顾的问题,提出了一种在潜在低秩分解基础上结合图像增强和显著性检测,并运用梯度导向滤波重构融合决策图的红外与可见光融合方法.首先使用增强算法提高可见光图像的细节轮廓清晰度,并对红外源图使用视觉显著性检测处理,提取最初的显著性权重图;接着对红外图像与增强图像进行潜在低秩分解,获取细节层和基础层,将细节层作为引导图像引入梯度导向滤波系统,对之前获取的显著性权重图进行优化,得到对图像细节和轮廓把握更加精准的二次权重图;然后将初次权重和二次权重作为融合决策图分别引导基础层和细节层的融合;最后将重构好的细节和基础层使用加权平均法进行融合得到最终结果.实验结果表明,算法对融合结果中细节信息的保留,边缘轮廓分辨度和红外目标显著性的提升,均有着较好的表现,在质量指标如平均梯度、视觉信息保真度、图像互信息等方面也取得了较好的效果.

To solve the insufficient background information in the fusion image and the problem that the infrared target salience and visible image detail preservation hard to achieve at the same time in the fusion process,in this paper,an infrared and visible light fusion method was proposed,which combined image enhancement and saliency detection on the basis of latent low rank decomposition,and used gradient guided filtering to reconstruct the fusion decision image.Firstly,the image enhancement algorithm was used to enhance the visible image to obtain a better performance of detail and contour,and the visual saliency detection algorithm(VSM) was used to processes the infrared image to obtain the first weight map.Secondly,the latent low-rank representation(LatLRR)decomposition was performed to obtain the detail layer and the base layer,the decomposed detail layer was taken as a parameter introduced into the gradient domain guidance filtering algorithm to filter the initial weight map,and got the improved weight map which preserving details and contour more accurately.Thirdly,the primary weight and secondary weight were used as the fusion strategy respectively to reconstruct the base layer and the detail layer.Finally,the weighted average algorithm was used to fuse the reconstructed details and base layer to obtain the final result.Experiments showed that the proposed method got better performance in preserving the background detail information,improving the infrared target’s salience and the identificational degree of contour features.It also got better results in quality indexs such as average gradient,visual information fidelity,mutual information and so on.