智能系统学报2026，Vol.21Issue(1)：95-108,14.DOI:10.11992/tis.202504002
基于多尺度协调卷积与自适应加权的红外与可见光图像融合
Infrared and visible image fusion based on multi-scale coordinated convolution and adaptive weighting
摘要
Abstract
To address the limitations of convolution neural networks-based image fusion models,such as restricted glob-al information perception,high-frequency detail preservation,and the loss function weights configuration,this article proposes a convolution and multilayer perceptron-integrated multiscale coordinate network(CM-MCNet)for high-qual-ity infrared and visible image fusion.In the encoder of CM-McNet,a convolutional weighted permute multilayer per-ceptron module is introduced to enhance spatial understanding by simulating feature permutation and integrates an ad-aptive feature reweighting mechanism to effectively capture global information.Meanwhile,a multiscale coordinate convolution(MsCConv)module is designed,leveraging the advantages of central difference convolution to enhance the retention and expression of high-frequency details.By incorporating multiscale parallel sub-networks,MsCConv en-sures the comprehensive preservation of multi-level features.Moreover,the embedded coordinate attention mechanism jointly modulates channel and spatial dimensions,enhancing complementary information while suppressing redundancy.Furthermore,a data-driven adaptive loss weighting strategy is proposed,which can dynamically adjust the contribution of supervision signals based on image feature statistics.This reduces the complexity of hyperparameter tuning while en-suring the loss function more accurately reflects the characteristics of the source images.Experimental results on the RoadScene,TNO,and M3FD public datasets demonstrate that CM-MCNet generates fused images with sharper edge preservation and more natural texture transitions.Additionally,our method achieves superior performance across vari-ous objective metrics,including information entropy,standard deviation,spatial frequency,visual information fidelity,and average gradient,outperforming existing state-of-the-art fusion methods.This work provides a novel perspective for infrared and visible image fusion and lays a solid foundation for further advancements in the field.关键词
图像融合/红外图像/可见光图像/多尺度协调卷积/卷积加权重排多层感知器/坐标注意力/自适应权重Key words
image fusion/infrared image/visible image/multiscale coordinate convolution/convolutional multilayer perceptron/coordinate attention/adaptive weighting分类
信息技术与安全科学引用本文复制引用
刘诗怡,刘金平,黄丽娟,蒋嘉豪,宋殿义,杨广益..基于多尺度协调卷积与自适应加权的红外与可见光图像融合[J].智能系统学报,2026,21(1):95-108,14.基金项目
国家自然科学基金项目(62371187) （62371187）
湖南省自然科学基金项目(2024JJ8309). （2024JJ8309）
