Registration and segmentation of PPL and XPL images of geological polished sections containing anisotropic minerals

We propose the neural network-based method for segmentation of minerals in images of geological polished sections. We use a specific image set where the images are taken in plane-polarized (PPL) and cross-polarized (XPL) light, with different rotation angles relative to the optical axis of the camera. The data set, formed in that way, allows to obtain additional information that improves the quality of the anisotropic mineral segmentation, as that type of minerals changes its color (“blinks“) depending on the rotation angle when imaging under the XPL light. The peculiarity of our method is the registration of the XPL images with the PPL images of the same polished section which is further fed to the neural network. Additionally, a data balancing algorithm was used to compensate for the non-uniform occurrence of different minerals in the image set. Five segmentation models have been trained both with using additional images and without using them. The results have demonstrated that using XPL images, registered in advance with the corresponding PPL image, improves the quality of segmentation by 3–12 percent for anisotropic minerals and by 1–8 percent for isotropic minerals.