Abstract: To address the reliance of existing unsupervised fusion networks on reconstruction constraints and static fusion modules, which prevents them from dynamically coordinating feature-processing paths according to bimodal content differences, a memory-gated expert network for infrared and visible image fusion was proposed. First, a dual-branch multi-scale feature encoding and expert fusion framework was constructed, in which a local spatial expert, a global channel expert, a structure-preserving expert, and a content-aware gating mechanism were used to dynamically coordinate different feature-processing paths. Then, a learnable prototype memory bank was introduced for each expert, and Top-k retrieval and soft-gated compensation were used to enhance degraded features. Finally, self-reconstruction, cross-reconstruction, gating-consistency, and memory regularization constraints were jointly imposed to improve the stability of unsupervised fusion. Experimental results show that, on MSRS, the Entropy (EN) and Structural Similarity Index Measure (SSIM) are 0.67% and 8.76% higher than those of the best-performing comparison network, SFDFuse and ITFuse, respectively. On LLVIP, EN and Standard Deviation (SD) are 0.84% and 7.95% higher than those of TDFusion, respectively. On TNO, SSIM is 3.58% higher than that of URFusion, while the Edge Information Fusion Quality Metric (Qabf) reaches 0.503 and ranks second. The network enhances infrared target saliency while preserving visible textures and structural information, thereby providing fused inputs that retain both target and background information for pedestrian detection in complex road scenes.

Key words: Keywords: image fusion, infrared and visible image fusion, unsupervised learning, gated expert network, memory compensation.

摘要： 摘 要: 针对现有无监督融合网络依赖重构约束与静态融合模块、难以依据双模态内容差异动态协调特征处理路径的问题，提出一种用于红外与可见光图像融合的记忆门控专家网络。首先，构建双分支多尺度特征编码与专家融合框架，通过空间局部、通道全局和结构保持专家及内容感知门控机制动态协调不同特征处理路径；其次，为各专家引入可学习原型记忆库，利用Top-k检索和软门控补偿增强退化特征；最后，联合自重构、交叉重构、门控一致性与记忆正则约束，提高无监督融合的稳定性。实验结果表明：在MSRS上，信息熵（EN）和结构相似性（SSIM）分别较最佳对比网络SFDFuse和ITFuse提高0.67%和8.76%；在LLVIP上，EN和标准差（SD）较TDFusion分别提高0.84%和7.95%；在TNO上，SSIM较URFusion提高3.58%，边缘信息融合质量指标（Qabf）达到0.503，取得次优结果。该网络能够在增强红外目标显著性的同时保持可见光纹理与结构信息，为复杂道路场景下的行人检测提供兼顾目标与背景信息的融合输入。

关键词: 图像融合, 红外与可见光图像融合, 无监督学习, 门控专家网络, 记忆补偿