基于MSHAM-UNet的岩心孔洞图像分割方法

Image Segmentation Method of Rock Core Hole Based on MSHAM-UNet

岩心图像的孔洞分割对于石油勘探有着重要意义。当前基于深度学习的孔洞分割方法存在着孔洞边缘分割不连续、分割精度低和参数量大等问题,为解决上述问题,提出一种基于UNet网络的改进模型MSHAM-UNet。首先,针对UNet模型对不同尺度特征图的跳级连接带来的语义信息丢失问题,设计一种结合HAM(hybrid attention module)的多尺度融合注意力模块(multi-scale hybrid fusion attention module, MSHAM),该模块对带有空间信息的浅层特征图和含有语义信息的深层特征图进行注意力特征融合,增强网络聚合不同尺度信息的能力。其次,使用GP-bneck模块替换部分普通卷积,在降低模型参数量和加深网络的同时,增强网络特征提取能力。实验结果表明,MSHAM-UNet网络在岩心孔洞数据集上的F1分数(F1-score)、交并比(intersection over union, IoU)和平均交并比(mean intersection over union, MIoU)分别达到了87.35%、77.27%和90.21%,相较于原始模型提高了5.29%%、4.02%和4.84%,对比主流的语义分割模型也有较高提升,为岩心孔隙研究提供了新的思路。

The hole segmentation of core images is important for oil exploration.In order to solve the problems of discontinuous hole edge segmentation,low segmentation accuracy and large number of parameters,an improved model based on UNet network MSHAMUNet was proposed.Firstly,aiming at the problem of semantic information loss caused by the hopping connection of UNet models to feature maps of different scales,a multi-scale hybrid fusion attention module(MSHAM)combined with a hybrid attention mechanism(HAM)was designed,which performed attention feature fusion on shallow feature maps with spatial information and deep feature maps with semantic information,so as to enhanced the ability of the network to aggregate information at different scales.Secondly,the GPbneck module was used to replace part of the ordinary convolution,which can reduce the number of model parameters and deepen the network,and enhance the network feature extraction ability.The experimental results show that the F1-score,intersection over union(IoU)and mean intersection over union(MIoU)of the MSHAM-UNet network on the core hole dataset reach 87.35%,77.27%and 90.21%,respectively,which are 5.29%,4.02%and 4.84%higher than the original model,and are also improved compared with the mainstream semantic segmentation model,which provides a new idea for core pore research.