强度图像和偏振度图像融合网络的设计  

Design of intensity image and polarization image fusion network

作　　者：闫德利 申冲 王晨光[2] 唐军 刘俊 YAN Deli;SHEN Chong;WANG Chenguang;TANG Jun;LIU Jun(School of Instrument and Electronics,North University of China,Taiyuan 030051,China;School of Information and Communication Engineering,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学仪器与电子学院,山西太原030051 [2]中北大学信息与通信工程学院,山西太原030051

出　　处：《光学精密工程》2023年第8期1256-1266,共11页Optics and Precision Engineering

基　　金：国家自然科学基金创新研究群体资助(No.51821003);国家自然科学基金优秀青年基金项目资助(No.51922009);国家自然科学基金面上项目资助(No.61973281);山西省重点研发计划项目资助(No.202003D111003);山西省优秀青年培育项目资助(No.202103021222011)。

摘　　要：为了弥补强度图像在阴暗处丢失纹理细节的劣势,结合偏振度图像的偏振特性,本文提出了一种强度图像和偏振度图像的融合方法。首先,构建编码器网络提取源图像的语义信息和纹理细节。随后,特征融合网络采用加法策略和残差网络进行图像特征融合。最后,通过解码器网络对融合后的图像特征进行重构获得最终的融合图像。此外,根据源图像和融合图像之间的结构相似性损失和梯度损失,本文提出了一种改进的损失函数来引导融合网络训练。实验结果表明:与其他6种方法中融合效果最好的改进的双通道脉冲耦合神经网络(MD-PCNN)相比,本文方法的客观评价指标平均梯度、信息熵、图像质量、标准差和改进的多尺度结构相似性分别提高了4.3%,1.0%,8.1%,2.5%,3.1%,图像噪声降低了8.8%,且克服了强度图像在阴暗处丢失纹理细节的问题。To overcome the disadvantage that the intensity image loses texture details in the dark,this study proposes a fusion method of the intensity and polarization images that combines the former with the polarization characteristics of the latter.First,an encoder network is constructed to extract the semantic in⁃formation and texture details of the source image.Subsequently,the feature fusion network adopts an addi⁃tive strategy and a residual network for image feature fusion.Finally,the fused image features are recon⁃structed through the decoder network to obtain the final fused image.Furthermore,according to the struc⁃tural similarity loss and gradient loss between the source and fused images,this study proposes an im⁃proved loss function to guide the fusion network training.Experimental results indicate that compared with the modified dual-channel pulse coupled neural network(MD-PCNN),which has the best fusion effect among the other six methods,the objective evaluation indicators of the proposed method—average gradi⁃ent,information entropy,image quality,standard deviation,and improved multi-scale structural similarity—are improved by 4.3%,1.0%,8.1%,2.5%,and 3.1%,respectively;the image noise is reduced by 8.8%.Moreover,the problem of losing texture details for intensity images is eliminated.

关 键 词：纹理细节 残差网络 结构相似性 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]