ISSN 1004-4140
CN 11-3017/P
王宸章, 王彦飞, 白治经. 基于伪标签方法的页岩孔隙语义分割网络[J]. CT理论与应用研究(中英文), xxxx, x(x): 1-12. DOI: 10.15953/j.ctta.2024.039.
引用本文: 王宸章, 王彦飞, 白治经. 基于伪标签方法的页岩孔隙语义分割网络[J]. CT理论与应用研究(中英文), xxxx, x(x): 1-12. DOI: 10.15953/j.ctta.2024.039.
WANG C Z, WANG Y F, BAI Z J. Shale-Pore Semantic Segmentation Network Based on Pseudo-Labeling[J]. CT Theory and Applications, xxxx, x(x): 1-12. DOI: 10.15953/j.ctta.2024.039. (in Chinese).
Citation: WANG C Z, WANG Y F, BAI Z J. Shale-Pore Semantic Segmentation Network Based on Pseudo-Labeling[J]. CT Theory and Applications, xxxx, x(x): 1-12. DOI: 10.15953/j.ctta.2024.039. (in Chinese).

基于伪标签方法的页岩孔隙语义分割网络

Shale-Pore Semantic Segmentation Network Based on Pseudo-Labeling

  • 摘要: 页岩孔隙是页岩气储集层的关键参数,其形状、大小、连通性和发育程度直接影响了储集性能。分析高分辨率岩石样本扫描电镜图像是获得孔隙结构的重要途径,但自动化识别仍然具有较大的难度。本研究提出了一种基于伪标签方法的页岩孔隙语义分割网络,试图实现页岩孔隙的智能识别和分类。我们采用金字塔场景解析网络(PSPNet),对251张重庆龙马溪组页岩储层的扫描电镜图像进行训练。由于采用了伪标签生成策略,只对少量图像进行标注,并借助模型在未标注图像上的分割结果对模型进行迭代训练,可以降低图像标注的成本。此外,还通过集成学习来提高模型的准确性和泛化能力。经过迭代训练后的模型,平均交并比(MIoU)可以高于0.70。经对比实验,采用伪标签和集成学习可以使模型的MIoU提升约0.07。结果表明,伪标签方法在提升神经网络泛化能力的同时,改善过往的深度学习分割方法人工标注时间成本过高的缺陷,而集成学习可以稳定地提升模型的准确率。

     

    Abstract: Shale pore structures contribute significantly to shale gas reservoirs, with their shape, size, connectivity, and development directly affecting storage. To achieve intelligent recognition and classification of shale pores, this study proposes a shale-pore semantic segmentation network based on a pseudo-label method. A total of 251 scanning electron microscopy images of the Longmaxi Formation shale reservoir in Chongqing are used, and the Pyramid Scene Parsing Network is utilized for training. Additionally, pseudo-label generation is employed, which involves annotating only a few images and using the model’s segmentation results on unlabeled images for iterative training. Subsequently, ensemble learning is conducted to improve the model’s accuracy and generalizability. The iteratively trained model has a mean intersection-over-union (MIoU) score exceeding 0.70. Comparative experiments show that employing pseudo-labeling and ensemble learning increases the model’s MIoU by approximately 0.07. Furthermore, pseudo-labeling enhances the generalizability of neural networks while addressing the high time cost of manual annotation required in previous deep-learning segmentation methods, whereas ensemble learning stably increases the model’s accuracy.

     

/

返回文章
返回