5:15 PM - 7:15 PM
[SCG60-P13] Development of automated detection methods for intragranular cracks using machine learning
Keywords:machine learning, Image processing, intragranular cracks
Faults exhibit different permeability from the host rock. A fault zone comprises a fault core, where the rock is highly comminuted, a surrounding damage zone characterized by an extensive crack network, and the host rock, which remains largely unaffected by faulting. While the fault core generally exhibits low permeability, the damage zone can display relatively high permeability due to its dense crack network. Therefore, a comprehensive understanding of fault permeability necessitates a detailed investigation of the internal structure of fault zones, particularly the characteristics of crack networks within the damage zone. One key factor affecting fault zone structure, especially crack network in the damage zone, is the crack distribution in the host rock. To investigate how this factor affects the induced crack network, Kikuchi et al. (2021) conducted axial deformation experiments on sandstone under confining pressure and analyzed pre- and post-deformation crack distributions. To evaluate the crack distribution, Kikuchi et al. (2021) manually detected densities of intragranular cracks in quartz, but this method is subjective, labor-intensive, and lacks consistency (Yamane & Chun, 2019). To address these issues, this study attempts to adapt image processing and machine learning. Previous studies on crack detections using image processing and machine learning mainly targeted cracks in rock surfaces or artificial structures (e.g. Gharineiat et al., 2022), whereas this study focuses on intragranular cracks, of which detection requires precise differentiation from grain boundaries. Few studies have specifically addressed intra-grain cracks, and established methods are limited. This study aims to develop an automated detection method for intra-grain cracks in quartz from thin-section images of sandstone using image processing and machine learning.
The pre-deformation thin sections of sandstone from Rajasthan, India, previously analyzed by Kikuchi et al. (2021), were used for the analyses (quartz 68.2%, plagioclase 9.5%, orthoclase 18.2%, mica 0.7%, others 3.3%; data from Kochi Core Research Institute). The mineral grain size is approximately 0.02–0.1 mm. This study employs backscattered electron (BSE) images and secondary electron (SE) images obtained by scanning electron microscope (SEM), and cross-polarized optical microscope images to detect intra-grain cracks in quartz. The detection process involves the following steps. Step 1: adjustment of orientation, position and size of images. Images of the same subject obtained using SEM and optical microscope were aligned using affine transformation, of which matrix are estimated to match points in each image corresponding to each other. Step 2: detection of quartz regions. SEM-BSE images were segmented into regions corresponding to quartz, non-quartz, resin and opaque minerals via machine learning using Trainable Weka Segmentation (TWS) of Fiji. Step 3: detection of quartz grain boundaries. To enhance quartz grain boundaries, HSV color space conversion, thresholding, binarization, and edge detection were applied to cross-polarized optical microscope images. Step 4: detection of cracks and distinction from grain boundaries. Morphological operations were applied to SEM-SE images to detect cracks and grain boundaries, and the grain boundaries are eliminated by comparing with the results of Step 3.
In Step 2, the quartz detection achieved 79.25% accuracy and 81.17% precision. In Step 1, despite brightness and contrast differences among the images, reasonably accurate alignments were achieved by selecting suitable corresponding points. The errors on the alignment depended on selected points, and the appropriate selection reduced the errors. For Steps 3 and 4, crack detections were tested using simulated images instead of actual SEM and polarized microscope images. Some grain boundaries were misidentified as cracks, possibly due to tonal information loss in Step 3 during binarization.
The pre-deformation thin sections of sandstone from Rajasthan, India, previously analyzed by Kikuchi et al. (2021), were used for the analyses (quartz 68.2%, plagioclase 9.5%, orthoclase 18.2%, mica 0.7%, others 3.3%; data from Kochi Core Research Institute). The mineral grain size is approximately 0.02–0.1 mm. This study employs backscattered electron (BSE) images and secondary electron (SE) images obtained by scanning electron microscope (SEM), and cross-polarized optical microscope images to detect intra-grain cracks in quartz. The detection process involves the following steps. Step 1: adjustment of orientation, position and size of images. Images of the same subject obtained using SEM and optical microscope were aligned using affine transformation, of which matrix are estimated to match points in each image corresponding to each other. Step 2: detection of quartz regions. SEM-BSE images were segmented into regions corresponding to quartz, non-quartz, resin and opaque minerals via machine learning using Trainable Weka Segmentation (TWS) of Fiji. Step 3: detection of quartz grain boundaries. To enhance quartz grain boundaries, HSV color space conversion, thresholding, binarization, and edge detection were applied to cross-polarized optical microscope images. Step 4: detection of cracks and distinction from grain boundaries. Morphological operations were applied to SEM-SE images to detect cracks and grain boundaries, and the grain boundaries are eliminated by comparing with the results of Step 3.
In Step 2, the quartz detection achieved 79.25% accuracy and 81.17% precision. In Step 1, despite brightness and contrast differences among the images, reasonably accurate alignments were achieved by selecting suitable corresponding points. The errors on the alignment depended on selected points, and the appropriate selection reduced the errors. For Steps 3 and 4, crack detections were tested using simulated images instead of actual SEM and polarized microscope images. Some grain boundaries were misidentified as cracks, possibly due to tonal information loss in Step 3 during binarization.