56 3. METHODOLOGY The identification and measurement of minerals on rock thin sections images can be automated using a deep learning workflow. Figure 1 illustrates the methodology, which consists of two main stages. The first stage focuses on data handling, while the second involves the AI model development. Figure 1 – Workflow for mineral identification and measurement using Deep Learning 3.1. Dataset creation Deep Learning relies on data. Several open-access online sources provide images of rock thin sections. Paucar et al. (2025) accessed platforms such as Virtual Microscope-The Open University (s.f.) and Britrocks-British Geological Survey (BGS) (2025), which host large image repositories. Using ShareX software, 80 images were collected, 40 in PPL and 40 in XPL, from five rock types: diorite, gabbro, granite, granodiorite, and syenite. In total, 400 images of 960 × 600 px in PNG (RGB) format were collected for the training and validation sets, along with an additional set of 20 images for evaluation. 3.2. Dataset preparation The BGS images present an overlap issue between PPL and XPL. To address this, the images were cropped so that all have dimensions of 900 × 540 px and were exported in JPG (RGB) format to reduce file size. This process was carried out using the trial version of Photoshop. The minerals were labeled using Roboflow, a platform that also allows the dataset to be split into training, validation, and evaluation subsets, as well as the application of data augmentation techniques (flipping, rotation, shearing, etc.). Eight minerals were labeled: K-feldspar (Afs), amphibole (Amp), biotite (Bt), muscovite (Ms), olivine (Ol), pyroxene (Px), plagioclase (Pl), and quartz (Qz), following the abbreviations of Whitney & Evans (2010). Figure 2 shows representative examples and the number of labels per mineral. Labels created on XPL images were duplicated onto the corresponding PPL images using the “Repeat Previous” option. Considering both XPL and PPL images, a total of 62178 minerals were labeled and exported as a JSON file. 3.3. Mineral identification and measurement The model consists of two components (Figure 3): the first identifies eight mineral categories using Detectron2, while the second performs object measurements using scikit-image. Matplotlib is used for histogram visualization, Python-Ternary for ternary diagrams, and FPDF for automated report generation. 3.3.1 Detectron 2-v0.6 (2021) It is a deep learning model developed by Facebook AI Research and implemented in PyTorch. It provides high-quality code for the implementation and evaluation of object detection and segmentation research (Girshick, 2018). Its three modules are described below. Backbone Network and Feature Pyramid Network (FPN): It is based on a CNN (ResNetX101) and extracts feature maps from the input image at multiple scales using the FPN. The convolutional layers (res2res5) reduce the image resolution, producing multiscale feature maps (P2-P6). Lateral convolutions adjust the number of channels to 256, and upsampling fuses low- and high-resolution information.
RkJQdWJsaXNoZXIy MTM0Mzk2