High-precision algorithm for grain segmentation of thin section by multi-angle optical-microscopic images

Abstract

The present paper introduces an algorithm for automating the analysis of quartz sandstones and siltstones petrographic thin sections images. The images of thin sections are obtained in polarized light at magnifications providing good image quality. In addition, the images for each section are obtained at different angles of rotation of the microscope stage. Augmentation is applied to the obtained photographs: the number of images increases due to rotations, shifts and rescaling of the image. For training the neural network of the Mask R-CNN architecture, transfer learning is used, with initial weights obtained from a huge variety of non-geological images. The results of image segmentation using Mask R-CNN are compared versus the Watershed algorithm results and the U-Net network for two metrics. According to the standard Intersection over Union metric, U-Net for high quality images and Watershed for blurry images show the best results with a slight superiority. However, according to the Grain Size Metri c, which evaluates the accuracy of grain size measurement, the best accuracy (over 95%) is shown by Mask R-CNN. The grain size analysis is done, and the porosity of the studied petrographic sections is determined. The use of the proposed approaches in the study of thin sections will significantly reduce the time for obtaining the results of grain size distribution analysis and porosity determination.