58 (loss_rpn_loc), applied only to detected objects. The top 1000 anchors are selected, and redundancies are removed using the Non-Maximum Suppression (NMS) algorithm. Region of Interest (ROI) Heads: This module receives the feature maps (P2-P6) and the proposals generated by the RPN. An ROI Pooler extracts features from each region and converts them into a onedimensional vector. Mask R-CNN uses two heads: the Box Head employs fully connected (FC) layers for object classification (loss_cls) and bounding-box regression to refine the proposals (loss_box_reg), the Mask Head performs instance segmentation using a fully convolutional network (FCN) that produces pixel-level masks, evaluated using binary cross-entropy (loss_mask). Finally, NMS removes redundant detections, yielding 100 boxes and masks by default; in this study, this number was increased to 600 per inference. The loss values from the RPN and ROI Heads modules are combined according to Equation 1. total_loss = loss_rpn_cls + loss_rpn_loc + loss_cls + loss_box_reg + loss_mask (1) 3.3.2 Image processing with scikit-image v0.25.2 (2025) Scikit-image is a Python library dedicated to image processing. Its aim is to provide a diverse collection of powerful and free tools. Its algorithms are widely used in research, education, and industrial applications (van der Walt, et al., 2014). The measure submodule and the regionprops function are used to extract measurements from each identified mineral, including contours, centroid, area, perimeter, and major and minor axis diameters. Pandas and NumPy are used to compute modal percentages, which are represented in the Streckeisen (1976) ternary diagram using Python-Ternary. Grain-size distributions are analyzed using histograms generated with Matplotlib, and the results are integrated into a PDF report. 4. EXPERIMENTATION AND RESULTS The aim of this section is to examine the Detectron2-v0.6 metrics for the training, validation, and evaluation datasets, as well as the results of mineral measurements obtained using scikit-image. 4.1. Computational platform Deep Learning models require a computational environment that integrates various hardware and software tools. Google Colaboratory is one of the most widely used platforms for this purpose. Table 1 summarizes the characteristics of the computational platform employed. Table 1 – Characteristics of the Google Colab computational platform Feature Specification Feature Specification Processor Intel ® Xeon ® Storage 235.7 GB CPU Speed 2.20 GHz Graphics Card (GPU) NVIDIA A100-SXM4-40GB CPU Cores 12 Python Version 3.11.11 Architecture x86_64 Compiler (GCC) 11.4.0 System RAM 83 GB Operating System Linux Ubuntu 22.04.3 LTS 4.2. Detectron2-v0.6 model metrics 4.2.1 Training dataset The 400 images were divided into 80% for training and 20% for validation. Using random combinations of flipping, 90° rotation, zoom, brightness, blurring, and noise, data augmentation was applied
RkJQdWJsaXNoZXIy MTM0Mzk2