Figure 3 – Mineral identification and measurement model: Detectron2 (Honda, 2020) and scikitimage. Region Proposal Network (RPN): Composed of three CNNs, it generates regions of interest with potential object presence. Anchors of different sizes are placed at each pixel of the P2-P6 feature maps. Two loss functions are used: objectness loss (loss_rpn_cls), which evaluates object presence, and localization loss (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 one-dimensional 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 crossentropy (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
RkJQdWJsaXNoZXIy MTM0Mzk2