基于多尺度跨阶段密集连接的图像融合算法  

Image fusion algorithm based on multi⁃scale cross⁃stage dense connection

作　　者：翟丽红[1] 罗继阳 ZHAI Lihong;LUO Jiyang(Taiyuan Institute of Technology,Taiyuan 030008,China;North Automatic Control Technology Institute,Taiyuan 030006,China)

机构地区：[1]太原工业学院,山西太原030008 [2]北方自动控制技术研究所,山西太原030006

出　　处：《现代电子技术》2025年第5期107-114,共8页Modern Electronics Technique

基　　金：山西省高等学校科技创新项目(2020L0671);2023年山西省高等学校一般性教学改革创新项目(J20231302)。

摘　　要：针对目前可见光与红外光图像融合过程中的关键细节信息丢失,目标对比度较低的问题,提出一种基于多尺度跨阶段密集连接网络的图像融合算法。通过多尺度卷积与跨阶段的密集连接网络实现双模态图像的特征提取工作,结合CA注意力机制提高模型的融合效果,并以L1范数作为特征融合规则来获取融合特征图,并最终通过解码网络实现图像的重构工作。实验结果表明,在公共数据集TNO中,文中提出的算法在结构相似度、信息熵以及差异相关系数三项指标中获得了最优值,相较于次优值分别提高了4.14%、2.66%、2.59%,在边缘信息度量上取得了次优值,与最优值相差3.3%。综合主客观评价,文中提出的方法可获取高质量的融合图像,具有明显的优势。In view of the key detail information loss and low contrast of the objects in the fusion process of visible and infrared images,an image fusion algorithm based on multi-scale cross-stage densely-connected network is proposed.The feature extraction of bimodal images is realized by multi-scale convolution and cross-stage densely-connected network,and the fusion effect of the model is improved by combining CA attention mechanism.The fusion feature map is obtained by taking the norm L1 as the feature fusion rule.Finally,the image is reconstructed by decoding the network.The experimental results show that in the public dataset TNO,the proposed algorithm achieves the optimal value for the three indicators of structural similarity,information entropy and difference correlation coefficient,which are 4.14%,2.66%and 2.59%higher than the sub-optimal value,and achieves the sub-optimal value for the edge information measurement,which is 3.3%lower than the optimal value.From the subjective and objective evaluation,it can be seen that the proposed method can obtain high-quality fusion images,and has obvious advantages.

关 键 词：图像处理 可见光与红外光 深度学习 图像融合 多尺度 跨阶段密集连接 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]