%0 Journal Article
%T 提升地震相边缘和不规则体识别效果的半监督识别方法<br>A Semi-Supervised Recognition Method to Improve the Recognition of Earthquake Phase Boundaries and Irregular Bodies
%A 李晓玲
%A 柴变芳
%A 硕良勋
%A 赵虹阳
%J Computer Science and Application
%P 1704-1712
%@ 2161-881X
%D 2023
%I Hans Publishing
%R 10.12677/CSA.2023.139168
%X <span style=\"font-size:14px;\">近年来随着深度学习的快速发展和地震相识别智能化研究深入，国内外提出了多种基于深度学习的地震相识别方法。但深度学习方法需要大量标注信息学习网络参数，实际上地震数据标注成本高，困难大。鉴于此，提出半监督地震相识别方法。利用少部分标注数据训练基于语义分割网络的地震相识别模型，并用模型对未标注的数据进行类别推断。该方法可扩展训练数据集规模，降低标注成本，提高模型的准确性和泛化能力。相较于仅用有限标注数据的模型性能有显著提升。该模型采用Deeplabv3+网络，在复杂深层相带存在边缘不清晰以及类内预测不一致问题，针对该问题，提出提升地震相边缘和不规则体识别效果的半监督识别方法。该方法在Deeplabv3+编码部分引入双注意力机制改善同类预测不一致现象，同时解码部分加入了有效通道注意力模块，结合低层特征进一步改善预测目标边界不清晰问题。改进后的Deeplabv3+网络有效改善存在的问题。</span><br />
<span style=\"font-size:14px;\">In recent years, with the rapid development of deep learning and indepth research on seismic phase recognition intelligence, various deep learning-based seismic phase recognition methods have been proposed both domestically and internationally. However, deep learning methods require a large amount of labeled information to learn network parameters, and in fact, the cost of seismic data labeling is high and difficult. In view of this, a semi-supervised seismic phase recognition method is proposed. The seismic phase recognition model based on a semantic segmentation network is trained using a small number of labeled data, and the model is used to infer categories for unlabeled data. This method can expand the training dataset scale, reduce the labeling cost, and improve the accuracy and generalization ability of the model. Compared with models that use only limited labeled data, the performance of this method has significant improvements. This model uses the Deeplabv3+ network, which has problems such as unclear edges in complex deep phase bands and inconsistent predictions within classes. To address these issues, a semi-supervised recognition method for improving the edge detection and irregular body recognition effect of seis-mic phases is proposed. This method introduces a double attention mechanism into the Deeplabv3+ encoding part to improve the problem of inconsistent predictions among similar classes, while adding an effective channel attention module to the decoding part to further improve the problem of unclear prediction targets by combining low-level features. The improved Deeplabv3+ network effectively improves the problem.</span>
%K 半监督学习，语义分割，地震相识别，有效通道注意力机制，双注意力机制<br>Semi-Supervised Learning
%K Semantic Segmentation
%K Seismic Facies Identification
%K Effective Channel Attention Mechanism
%K Dual Attention Mechanism
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=72428