Scholarlink

<p>The detection of rock fractures from formation micro-imager well-log images provides valuable information for systematic and objective characterization of subsurface formations. An evaluation is presented that compares the performance of two advanced automated image feature detection, deep-learning-based convolutional neural network (CNN) instance segmentation models: (YOLOv8n-seg), a one-stage model; and (Mask R-CNN), a two- stage model. To enhance the visibility and diversity of fracture-feature recognition, image inpainting and augmentation techniques are applied before fractures are annotated to provide a high quality labelled features for model training. The two models are trained and tested on the labeled dataset, with accuracy, recall, and inference speed used to quantify their fracture-recognition performance. The trained YOLOv8n-seg model attained a recall of 1.00 and a precision of 0.961 for all fracture-feature classes considered, while the Mask R-CNN attained a recall of 1.00 and a precision of 0.924 when applied to the testing subset. The inference times for YOLOv8n-seg model were 161 frames per second, whereas the Mask R-CNN model achieved only 27 frames per second (5.83 times slower). The results demonstrate that the single-stage (YOLOv8n-seg) model outperforms the two-stage model (Mask R-CNN) model for well-log image feature analysis. It does so in terms of accuracy and efficiency, and could be used for real-time, automated, fracture detection from well-log images in a variety of geological settings. A combined approach involving both algorithms offers the potential to improve fault and fracture detection from FMI imaging data, with possible extension to other subsurface data types, such as seismic data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A comparative study of YOLOv8 and Mask R-CNN for automated geological fracture detection in well-log images

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Muhammad Rashid,
  • Miao Luo,
  • David A. Wood

摘要

The detection of rock fractures from formation micro-imager well-log images provides valuable information for systematic and objective characterization of subsurface formations. An evaluation is presented that compares the performance of two advanced automated image feature detection, deep-learning-based convolutional neural network (CNN) instance segmentation models: (YOLOv8n-seg), a one-stage model; and (Mask R-CNN), a two- stage model. To enhance the visibility and diversity of fracture-feature recognition, image inpainting and augmentation techniques are applied before fractures are annotated to provide a high quality labelled features for model training. The two models are trained and tested on the labeled dataset, with accuracy, recall, and inference speed used to quantify their fracture-recognition performance. The trained YOLOv8n-seg model attained a recall of 1.00 and a precision of 0.961 for all fracture-feature classes considered, while the Mask R-CNN attained a recall of 1.00 and a precision of 0.924 when applied to the testing subset. The inference times for YOLOv8n-seg model were 161 frames per second, whereas the Mask R-CNN model achieved only 27 frames per second (5.83 times slower). The results demonstrate that the single-stage (YOLOv8n-seg) model outperforms the two-stage model (Mask R-CNN) model for well-log image feature analysis. It does so in terms of accuracy and efficiency, and could be used for real-time, automated, fracture detection from well-log images in a variety of geological settings. A combined approach involving both algorithms offers the potential to improve fault and fracture detection from FMI imaging data, with possible extension to other subsurface data types, such as seismic data.