DSEL-CNN:结合注意力机制与均衡损失的图像融合算法  

DSEL-CNN: Image Fusion Algorithm Combining Attention Mechanism and Balanced Loss

作　　者：赵雅婷 韩龙 何辉煌 陈楚 ZHAO Yating;HAN Long;HE Huihuang;CHEN Chu(School of Electrical&Control Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)

机构地区：[1]黑龙江科技大学电气与控制工程学院,黑龙江哈尔滨150022

出　　处：《红外技术》2025年第3期358-366,共9页Infrared Technology

基　　金：黑龙江省重点研发计划项目(GA23A910);黑龙江省自然科学基金项目(LH2021F051);黑龙江科技大学大学生创新创业训练计划项目(YJS2021051)。

摘　　要：在红外与可见光图像融合时,融合后图像常出现显著目标不突出、可见光信息表达不充分的问题,且在亮度不均衡条件下,易出现边缘模糊和局部信息不均衡。因此,提出了结合注意力机制与均衡损失的图像融合算法(DepthwiseSeparable,Squeeze-and-Excitation,andEquilibriumLoss-based Convolutional Neural Network, DSEL-CNN)。首先,使用深度可分离卷积提取图像特征;其次,在融合策略中使用Squeeze-and-Excitation注意力机制来提高有效信息的权重;最后,利用均衡组合损失函数计算融合后图像损失,进行图像信息均衡。与FusionGAN、Dense Fuse和其它4种融合算法在TNO和MSRS公开数据集中进行主客观对比实验,其中互信息值、视觉信息保真度、边缘信息保留指标较其它6种算法分别最高提高了1.033、0.083、0.069,实验结果表明所提算法与6种常用融合算法相比,在融合图像视觉感观、信息含量、边缘和纹理保留方面均有提升。In infrared and visible image fusion,fused images often suffer from insufficient prominence of significant targets,inadequate expression of visible light information,edge blurring,and local information imbalance under uneven lighting conditions.To address these issues,an image fusion algorithm that combines attention mechanisms and equilibrium loss,termed the depthwise separable,squeeze-and-excitation,and equilibrium loss-based convolutional neural network(DSEL-CNN),is proposed.First,a depth-wise separable convolution is used to extract the image features.Subsequently,a fusion strategy is used to apply the squeeze-and-excitation attention mechanism to enhance the weight of effective information.Finally,an equilibrium composite loss function is utilized to calculate the loss of the fused image to ensure balanced information.A comparison of the fusion generative adversarial network(FusionGAN),DenseFuse,and four other fusion algorithms on the TNO and multi-spectral road scenarios(MSRS)public datasets showed that the proposed method achieved the highest improvements in mutual information(MI),visual information fidelity(VIF),and edge retention index(Qabf)by 1.033,0.083,and 0.069,respectively.Experimental results demonstrate that the proposed algorithm outperforms six commonly used fusion methods in terms of visual perception,information content,and edge and texture preservation in fused images.

关 键 词：图像处理 图像融合 深度可分离卷积 注意力机制 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]