基于像素差分神经网络的断层识别方法

Fault detection method based on pixel difference network

断层识别是地震解释的一项重要任务.相对传统的基于人工或半人工的断层检测方法,基于深度学习的断层检测方法具有自动化程度高等特点,近年来已经吸引了业界广泛兴趣.目前,大多数利用深度学习的断层识别方法都是基于语义分割或图像分类,但基于语义分割或图像分类的方法预测出的断层噪声较多.而边缘检测网络可以通过学习断层在地震剖面中不连续的特征,从而提高网络的抗噪能力.为了利用边缘检测网络的抗噪优点并提升其性能,本文采用基于像素差分的卷积运算构建出像素差分网络模型(Pixel Difference Networks,Pidinet).相较于传统的边缘检测网络,Pidinet将传统的边缘检测算子与深度学习网络结合,有效地提升了边缘检测的效果.为了使Pidinet网络更好地预测断层,本文对原始Pidinet网络进行了优化,去掉了相对断层识别任务而言多余的分支结构和卷积层.相对于传统的卷积运算,基于像素差分的卷积运算可以让神经网络更好地学习断层信息.为了让网络充分学习数据中的断层特征,本文将真实地震样本与合成地震样本混合得到训练所用数据集.实验证明Pidinet在交并比(Intersection over Union,IoU)方面较Holistically Nested Edge Detection(HED)边缘检测网络提升了10%左右.为了测试网络的迁移能力,本文只使用少量的数据样本对网络进行微调,迁移的结果较分类网络在F1分数(F1 Score)、灵敏度(sensitivity)等指标均提升了10%以上.最后,本文使用公开的实际地震数据进行测试,实验结果显示Pidinet识别出的断层连续且清晰,从而证明了基于边缘检测的深度学习算法在断层识别问题中的有效性.

Fault detection is an important task of seismic interpretation.Recently,deep learning methods have attracted huge attention in the industry as they have substantially improved automation compared to traditional manual or semi-manual based fault detection methods.Currently,most deep-leaning-based fault detection methods are based on semantic segmentation or image classification,but the faults predicted by these methods are noisy.By contrast,the edge detection network resists noise better by learning the discontinuous characteristics of faults in the seismic profile.In order to utilize the anti-noise advantage of edge-detection networks,as well as,improve their performance,in this paper,we use pixel-difference-based convolutional operators to construct a network model named Pixel Difference Networks—Pidinet.Compared with the traditional edge detection network,Pidinet combines the traditional edge detection operator with the deep learning network to effectively improve the effect of edge detection.For better predicting faults,we further optimized the original Pidinet by removing some unnecessary branches and convolution layers.The pixel-difference-based convolution operator boosts the neural network to learn fault information better than the traditional convolution operators.For fully learning the fault features in the data,a small amount of real seismic data samples is mixed with the synthetic seismic data samples in the training sample set.Tests on both synthetic and real data sets demonstrated that Pidinet improved the Intersection over Union(IoU)by about 10%compared to Holistically Nested Edge Detection(HED)network.Furthermore,to test the transfer-learning ability of the network,only a small amount of data is used to further fine-tune the network.Compared to the classification network,the transfer learning results improved by more than 10%in terms of F1 score and sensitivity.Finally,the publicly available real seismic data are used for the test.The experimental results show that the faults identified by Pidinet are continuous and clear,thus demonstrating the effectiveness of the edge-detection-based deep-learning algorithms for fault detection.