基于U-Net网络的成像测井图像修复方法研究  

Research on Imaging Logging Image Restoration Method Based on U-Net Network

作　　者：曹茂俊[1] 冯昊 CAO Mao-jun;FENG Hao(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《计算机技术与发展》2023年第6期61-68,共8页Computer Technology and Development

基　　金：黑龙江省自然科学基金(LH2019F004);东北石油大学优秀中青年科研创新团队(KYCXTD201903)。

摘　　要：针对微电阻率成像测井图像部分缺失的问题,该文提出了一种基于改进U-Net卷积神经网络的修复方法。该方法使用了VGG16网络的预训练权重,对模型参数进行初始化;引入混合空洞卷积增大感受野来捕获多尺度缺失区域信息;并且通过在模型中引入双向注意力图模块,前向注意力图进行缺失区域权重的重加权,反向注意力图注重修复区域的质量提升。根据实验结果,测试集中五组缺失区域大小不同的成像测井图像的平均结构相似性度量为0.93,相比其他同类方法提升了0.25左右。研究表明,该方法可用于微电阻率成像测井图像的修复,并在语义结构连贯、纹理细节等方面有不错的提升,从而为保证对微电阻率成像测井图像后续解释的顺利推进提供了一种新的深度学习方法。Aiming at the problem of missing parts of micro-resistivity imaging logging images,we propose a repair method based on an improved U-Net convolutional neural network.The method uses the pre-training weights of the VGG16 network to initialize the model parameters.The mixed hole convolution is introduced to increase the receptive field to capture multi-scale missing area information.By introducing a bidirectional attention map module into the model,the forward attention map reweights the missing area,and the reverse attention map focuses on the quality improvement of the repair area.According to the experimental results,the average structural similarity measure of five groups of imaging logging images with different missing area sizes in the test set is 0.93,which is about 0.25 higher than other similar methods.The research shows that the proposed method can be used for the restoration of micro-resistivity imaging logging images,and has a good improvement in semantic structure coherence and texture details,which provides a new deep learning method to ensure the smooth promotion of the subsequent interpretation of micro-resistivity imaging logging images.

关 键 词：微电阻率成像测井 U-Net卷积神经网络 VGG16网络 混合空洞卷积 注意力图 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]