论文详情
成像测井解释中裂缝解释是油气田开发地质研究的重要内容之一, 然而裂缝分割一直以来是成像测井解释的难点之一。为此, 提出一种新的融合了注意力机制的测井图像裂缝分割方法。首先, 在UNet分割网络结构的启发下, 构造一种基于跳跃连接的卷积神经网络模型, 提升了编码和解码部分的信息交换和信息融合, 以提取裂缝丰富的结构信息。然后, 为了对裂缝进行准确分割, 在网络解码部分中的连接层后面加入了注意力机制模块, 提升了网络捕捉裂缝全局信息的能力。最后, 利用最小生成树算法修复了存在断点的裂缝, 有利于更好地统计裂缝信息, 实现后续的裂缝定量分析。相较现有的图像裂缝分割方法, 该方法的分割结果无论在主观视觉上还是客观指标评价上都是最优的。实验结果表明, 所提出的方法能够准确地提取出测井图像的裂缝信息, 为后续裂缝参数的定量计算及测井资料解释奠定了良好基础, 具有较好的实用性。
Fracture segmentation still remains to be solved as the detection and characterization of fractures via image log interpretation are essential to oil and gas development and geological research.In this study, a novel method of fracture segmentation based on attention mechanisms is proposed.First, inspired by the structure of a UNet segmentation network, a convolutional neural network based on a jump connection was constructed to improve the information exchange and fusion of encoding and decoding in order to extract the structurally rich information of fractures.Second, to extract information concerning fractures accurately, the attention mechanism module was introduced after the connection layer in the decoding part of the network that improves the ability of the network to capture the global information of fractures.Finally, the minimum spanning tree algorithm was used to inpaint fractures with breakpoints, thus characterizing fractures better and realizing the subsequent quantitative analysis.Compared to the existing segmentation methods, the novel segmentation method developed in this study achieved the best performance in terms of a subjective vision evaluation and an objective index evaluation.The experimental results verify the practicability and accuracy of this method in the information extraction of fractures which in turn pave the way for the subsequent quantification of fracture parameters as well as image log interpretation.