双分支动态感受野的SAR与可见光融合算法  

Dual-branch Dynamic Receptive Field Fusion Algorithm of SAR and Visible Light

作　　者：王程远 刘立群[1] WANG Chengyuan;LIU Liqun(Information Science and Technology,Gansu Agricultural University,Lanzhou 730070,China)

机构地区：[1]甘肃农业大学信息科学技术学院,甘肃兰州730070

出　　处：《软件工程》2025年第4期43-46,52,共5页Software Engineering

基　　金：国家自然科学基金(32460440);甘肃省高校教师创新基金项目(2023A-051)。

摘　　要：针对合成孔径雷达(SAR)与可见光图像融合算法中特征提取不全、纹理细节丢失等问题,提出一种双分支动态感受野的融合算法。首先,利用双分支分别提取SAR和可见光图像特征。其次,针对可见光分支引入可学习的权重选择单元指导融合。最后,引入了结构相似性损失与改进的峰值信噪比损失函数,更好地保留了SAR和可见光的结构纹理信息。实验结果显示,该算法在主观视觉方面表现出色,细节信息丰富且人眼辨识度高,同时在客观评价指标(包括GD、SF、PSNR、SSIM及STD)上也实现了显著的提升。Aiming at the issues of incomplete feature extraction and loss of texture details in the fusion algorithms of Synthetic Aperture Radar(SAR)and visible light images,a dual-branch dynamic receptive field fusion algorithm is proposed.Firstly,the dual branches are utilized to separately extract features from SAR and visible light images.Secondly,a learnable weight selection unit is introduced into the visible light branch to guide the fusion process.Lastly,a structural similarity loss and an improved Peak Signal-to-Noise Ratio(PSNR)loss function are incorporated to better preserve the structural and textural information of both SAR and visible light images.Experimental results demonstrate that the proposed algorithm excels in subjective visual performance,with rich detail information and high human eye discernibility.Additionally,it achieves significant improvements in objective evaluation metrics,including Gradient Difference(GD),Spatial Frequency(SF),PSNR,Structural Similarity Index(SSIM),and Standard Deviation(STD).

关 键 词：条件生成对抗网络 双分支 动态感受野 残差网络 图像融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]