基于深度学习的偏振图像融合方法  

Deep Learning-Based Polarization Image Fusion Method

作　　者：孙红雨 李军 袁博 周宇超 SUN Hongyu;LI Jun;YUAN Bo;ZHOU Yuchao(Institute of Electronic and Information Engineering,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]山东科技大学电子信息工程学院,山东青岛266590

出　　处：《红外技术》2025年第2期193-200,共8页Infrared Technology

基　　金：山东省自然科学基金(ZR2024QF085)。

摘　　要：为改善阴暗复杂环境下图像质量,综合利用偏振图像的全局信息和纹理细节,提出多尺度特征提取与双重融合策略网络(scale feature extraction and dual fusion strategy, SFE-DFS-Nest),用于强度图像与偏振度图像的融合。首先,构建编码器,实现源图像多尺度特征提取。其次,浅层特征通过轻量化Transformer融合,深层特征通过残差网络融合。最后,构建解码器,用于融合特征重建。与现有图像融合网络相比,该网络针对不同尺度特征采用不同融合策略。结果表明,经过该网络融合后的阴暗复杂环境图像,主观视觉上图像观察舒适性较佳。并且通过选取方法的对比,融合后的图像在客观评价指标上,皆优于选取的方法。To improve image quality in complex and dim environments,a network that leverages both the global information and textural details of polarized images through a strategy of multi-scale feature extraction and dual fusion,known as the Scale Feature Extraction and Dual Fusion Strategy Network(SFE-DFS-Nest),is proposed.The proposed network fuses polarized intensity images with polarization degree images.Initially,an encoder is constructed to extract multi-scale features from source images.Then,shallow features are fused using a lightweight Transformer,while deep features are integrated through a residual network.Finally,a decoder is built to reconstruct the fused features.Compared with existing image fusion networks,this network employs distinct fusion strategies for features at different scales.The experimental results show that images from dark and complex environments exhibited improved subjective visual comfort after fusion through this network.Furthermore,the fused images obtained using the proposed method outperformed those obtained using the compared methods in terms of objective evaluation metrics.

关 键 词：SFE-DFS-Nest网络 偏振图像 图像融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]