5:15 PM - 6:45 PM
[MGI30-P01] Development of automated detection methods for quartz grains using deep learning
Keywords:Deep learning, quartz grains, Auto detection, Confusion matrix
Faults developed in rock have a permeability different from undeformed host rock. Permeability along sufaults strongly depends on the internal structure of the fault zones, in particular the degree of development of the fracture network. The internal structure of a fault zone is influenced by a variety of factors. One of these factors is the distribution of fractures within the rocks prior to fault formation. With the aim of investigating the influence of the initial distribution of fractures in the host rock on the internal structure of the fault, Kikuchi (2021) carried out axial deformation experiments with Indian sandstone specimens under confining pressure and the internal structure of the specimens was analyzed before and after deformation. The analysis of the internal structure involved the derivation of the intra-grain fracture density as a function of distance from the center of the fault. Only quartz grains were used to derive the intra-grain fracture density, in order to avoid counting the pre-deformation fractures as intra-grain fractures. Kikuchi (2021) detected quartz grains manually. There are problems with this method, such that the results might depend on the ability and experience of the detector, highly accurate and objective evaluation is difficult, and the enormous amount of time and labour is required (Yamane and Chun, 2019). The aim of this study was to develop a method for automatic detection of quartz grains from Indian sandstone this section images using machine learning and image processing.
The sample for experiment is Indian sandstone (quartz 68.2%, plagioclase 9.5%, potassium feldspar 18.2%, mica 0.7%, others 3.3%; information provided by the Kochi Core Research Institute) . A thin section of a pre-deformed sample of the rock was used. The grain size of the minerals in this rock is approximately 0.02 - 0.1 mm. Trainable Weka Segmentation (TWS; Arganda-Carras et al., 2017), a plug-in to the image processing package Fiji, was used to automatically detect quartz grains. The performance of machine learning was evaluated by selecting different teacher data set from within the analyzed images. The analyzed images used for machine learning were obtained using a scanning electron microscope (SEM) in reflection electron-composition image mode, incident voltage: 15.0 kV, WD: 13.2 mm and low vacuum. The resolution was 1 pixel = 10.8 μm. Each analyzed image was binarized and converted to numerical data using MATLAB, the images of quartz grains manually detected from the SEM images (correct images) were compared with the images from the attempted automatic detection using TWS, Accuracy and Precision were then calculated from the confusion matrix.
The maximum Accuracy and Precision was 78.68% and 80.74 %, respectively. However, results also showed the case that the addition of teacher data caused over-learning and reduced the accuracy of the machine learning. Therefore, it was found that the accuracy of machine learning depends on the type, not the number, of teacher data.
In this study, a method for automatic detection of quartz grains from sandstone this section images were developed using a combination of machine learning and image processing. Results showed improved accuracy in some trials but suggested that this was influenced by over-learning and noise. Future research will aim to improve this method by processing noise in analyzed images and expanding the TWS algorithm and dataset, including automatic detection of intra-grain cracks in quartz.
I. Arganda-Carreras., A. Cardona., V. Kaynig., J. Schindelin (2017) Bioinformatics, 33 (15), pp. 2424-2426.
Yuka Kikuchi1 , Shinichi Uehara, Kazuo Mizoguchi (2021): Japan Geoscience Union Meeting 2021. Effect of initial crack distribution on inner structure and permeability of fault zone
Tatsuro Yamane, Pang-jo Chun(2019):Crack detection from an image of concrete surface based on semantic segmentation by deep learning, Journal of JSCE, Vol.65A, pp.130-138.
The sample for experiment is Indian sandstone (quartz 68.2%, plagioclase 9.5%, potassium feldspar 18.2%, mica 0.7%, others 3.3%; information provided by the Kochi Core Research Institute) . A thin section of a pre-deformed sample of the rock was used. The grain size of the minerals in this rock is approximately 0.02 - 0.1 mm. Trainable Weka Segmentation (TWS; Arganda-Carras et al., 2017), a plug-in to the image processing package Fiji, was used to automatically detect quartz grains. The performance of machine learning was evaluated by selecting different teacher data set from within the analyzed images. The analyzed images used for machine learning were obtained using a scanning electron microscope (SEM) in reflection electron-composition image mode, incident voltage: 15.0 kV, WD: 13.2 mm and low vacuum. The resolution was 1 pixel = 10.8 μm. Each analyzed image was binarized and converted to numerical data using MATLAB, the images of quartz grains manually detected from the SEM images (correct images) were compared with the images from the attempted automatic detection using TWS, Accuracy and Precision were then calculated from the confusion matrix.
The maximum Accuracy and Precision was 78.68% and 80.74 %, respectively. However, results also showed the case that the addition of teacher data caused over-learning and reduced the accuracy of the machine learning. Therefore, it was found that the accuracy of machine learning depends on the type, not the number, of teacher data.
In this study, a method for automatic detection of quartz grains from sandstone this section images were developed using a combination of machine learning and image processing. Results showed improved accuracy in some trials but suggested that this was influenced by over-learning and noise. Future research will aim to improve this method by processing noise in analyzed images and expanding the TWS algorithm and dataset, including automatic detection of intra-grain cracks in quartz.
I. Arganda-Carreras., A. Cardona., V. Kaynig., J. Schindelin (2017) Bioinformatics, 33 (15), pp. 2424-2426.
Yuka Kikuchi1 , Shinichi Uehara, Kazuo Mizoguchi (2021): Japan Geoscience Union Meeting 2021. Effect of initial crack distribution on inner structure and permeability of fault zone
Tatsuro Yamane, Pang-jo Chun(2019):Crack detection from an image of concrete surface based on semantic segmentation by deep learning, Journal of JSCE, Vol.65A, pp.130-138.