基于特征掩膜优化的多聚焦图像融合方法  

Multi-Focus Image Fusion Method Based on Feature Mask Optimization

作　　者：王程 杨馨 刘晓文 WANG Cheng;YANG Xin;LIU Xiao-wen(College of Information and Cyber Security,People's Public Security University of China,Beijing 100038,China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100038

出　　处：《计算机仿真》2025年第2期226-234,共9页Computer Simulation

基　　金：国家重点研发计划项目(2022YFC31400);高分辨率对地观测系统重大专项(01-Y30F05-9001-20/22,GFZX0404130307)。

摘　　要：针对深度学习多聚焦图像融合算法中存在的特征表达不充分、纹理细节不完整和图像边缘模糊等问题,为了充分提取源图像的聚焦信息,提出了一种基于特征掩膜优化的多聚焦图像融合方法。针对深度学习网络中特征提取时的表达不充分,首先在网络编码阶段引人空间注意力模块对提取过程进行特征增强,深化聚焦区域与散焦区域间的差异性,得到可定位聚焦区域的掩膜图;其次利用阈值分割对编码网络得到的特征图进行二值化,实现图像中聚焦与散焦区域的分离并消除无关区域的影响;最后通过解码进行重建得到融合结果。将结果在Lytro数据集与MFFW数据集中与8种融合方法进行对比,并且分别从定性和定量分析两个方法评价融合结果。在定性分析上,上述方法在边缘清晰度、细节完整性、色彩鲜明度、对比度等方面均优于其它对比方法;在定量分析上,上述方法在Lytro数据集中结构相似性、峰值信噪比和均方根误差三种评价指标均值最优,相关系数评价指标均值次优,在MFFW数据集中相关系数、峰值信噪比、均方根误差三项指标均值最优。实验结果表明,融合方法较好地解决了深度学习中特征表达不充分、纹理细节不完整和图像模糊等问题,更好地保留了聚焦区域的特征信息和边缘区域的纹理细节,视觉效果显著。Aiming at the problems of insufficient feature expression,incomplete texture details,and blurred image edges in deep learning multi focus image fusion algorithms,a multi focus image fusion method based on feature mask optimization is proposed to fully extract the focus information of the source image.In response to the insufficient expression of feature extraction in deep learning networks,a spatial attention module is first introduced in the network encoding stage to enhance the feature extraction process,deepen the difference between the focused area and the defocused area,and obtain a mask map that can locate the focused area;Secondly,threshold segmentation is used to binarize the feature maps obtained by the encoding network,achieving the separation of focused and defocused regions in the image and eliminating the influence of irrelevant regions;Finally,the fusion result is obtained through decoding and reconstruction.Compare the results with 8 fusion methods on the Lytro dataset and MFFW dataset,and evaluate the fusion results using both qualitative and quantitative analysis methods.In qualitative analysis,this method is superior to other contrast methods in edge clarity,detail integrity,color freshness,contrast and others;In terms of quantitative analysis,the method in this paper has the best mean value of three evaluation indicators:structural similarity,peak signal to noise ratio and root mean square error in Lytro dataset,the second best mean value of correlation coefficient evaluation indicators,and the best mean value of correlation coefficient,peak signal to noise ratio and root mean square error in MFFW dataset.Conclusion The experimental results show that the proposed fusion method can solve the problems of feature loss,incomplete texture details and image blurring effectively.Meanwhile,it preserves the feature information of the focus areas and the texture details of the edge areas well.The visual effect is remarkable and it has adaptability and robustness.

关 键 词：深度学习 多聚焦图像 图像融合 有监督训练 掩膜 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]