基于深度学习的岩石薄片正交偏光图像多曝光融合算法  

Multi-exposure Fusion Algorithm for Orthogonal Polarized Images of Rock Section Based on Deep Learning

作　　者：王杰 晏鹏程 何小海[1] 滕奇志[1] Wang Jie;Yan Pengcheng;He Xiaohai;Teng Qizhi(Institute of Image Infonnation,College of Electronics and Infonnation Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院图像信息研究所,四川成都610065

出　　处：《长江信息通信》2021年第3期47-51,共5页Changjiang Information & Communications

摘　　要：针对正交偏光下岩石薄片图像中动态范围较低,无法观察全部颗粒的问题,将深度学习应用于多曝光图像融合算法,通过融合多个曝光度的薄片图像来获取较高的动态范围。首先将低动态范围的多曝光图像序列输入卷积神经网络,然后网络通过优化损失函数获取较好的权重图预测结果,最后由源多曝光图像序列联合权重图加权融合得到具有较高动态范围的薄片图像。对比实验表明,该算法可以有效地提升图像中的颗粒清晰度。Aiming at the problem that the dynamic range of rock section images under orthogonal polarized light is low,and all granules cannot be observed,deep learning is applied to the multi-exposure image fusion algorithm to obtain a higher dynamic range by fusing section images with multiple exposures.First,the low dynamic range multi-exposure image sequence is input into the convolutional neural network,and then the network obtains a better weight map prediction result by optimizing the loss function,and finally the source multi-exposure image sequence is weighted and fused with the weight map to obtain a higher dynamic range thin section image.Comparative experiments show that the algorithm in this paper can effectively improve the definition of particles in the image.

关 键 词：正交偏光图像 多曝光图像融合 混合空洞卷积 密集连接 神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]