基于全局双组注意力的红外与可见光图像融合

打开文本图片集

关键词：红外与可见光图像融合；图像增强；全局双组注意力；空洞卷积 中图分类号：TP391.4文献标识码：Adoi：1O.37188/CJLCD.2025-0195 CSTR：32172.14.CJLCD.2025-0195

Infrared and visible image fusion based on global dual-group attention

ZHAO Yang*，YANG Wengui， GAO Cuiyun （School of Electronic and Information Engineering，Anhui Jianzhu University， Hefei 230601，China）

Abstract： In complex scenarios，infrared and visible image fusion models often struggle to full extract the characteristics of overall macro-structures（from infrared images）and local micro-details （from visible images），as wellas to achieve synergy between these elements，which degrades fusion quality. To address this problem，this paper proposes a collaborative fusion principle based on scale specialization and designs a new fusion model based on an autoencoder architecture.The encoder and decoder of the model adopt a convolutional neural network （CNN） architecture. The model utilizes the global dual-group attention mechanism：after grouping feature maps by length and width to extract information，the generated inter-group channel attntion map can achieve weighting of the feature maps，thereby generating new feature maps containing more large-scale global structural information. The model utilises a convolution mechanism with multi-scale pooling and dilation，using receptive fields of different sizes and implementing global average and median pooling operations，to extract small-scale local features in the image. The model utilizes a decoder to integrate the large-scale structure and small-scale details of densely connected layers and skip connections，enabling them to synergistically fuse and reconstruct the fused image. The experimental results demonstrate that，on the MSRS and TNO datasets，compared to the best results of other methods， the information entropy，mean gradient，and edge intensity were improved by 0.95% ， 6.28% ，and 6.19% ， and then by 1.75% ， 13.51% ，and 11.75% respectively. Spatial frequency increased by 4.61% on the MSRS dataset，second only to the MDLSR-RFM method on the TNO dataset. These results validate the improvement in the quality of merged images in complex scenarios，as wellas the increased robustness and generalization of the model.

Key words： infrared and visible image fusion；image enhancement；global dual-group attention；dilated convolution

1引言

红外与可见光图像融合通过整合红外和可见光图像中的关键特征，设计合理的融合策略以实现信息互补，从而生成兼具多模态优势的合成图像[1-3]。（剩余15552字）