137 4.2. Model architecture The proposed architecture has two complementary outputs. First, a CNN-based semantic segmentation model produces a pixel-level lithological-domain map. A U-Net-type architecture is appropriate for this task because it combines local texture recognition with spatial localization through an encoder-decoder structure. Second, a YOLO-type detector identifies non-geological occlusions and operational objects that should be masked, down-weighted or reviewed separately. These include mesh, support components, water reflections, lamps, paint marks and fragmented material. The learning strategy is staged. A domain classifier is first trained on image tiles to verify that the selected classes are visually separable. The segmentation model is then trained on geologist-drawn polygons in representative areas. The object detector is trained with bounding boxes for occlusions and operational objects. Finally, both outputs are combined into a reviewer-facing map composed of a lithological-domain layer, an object or occlusion layer and a confidence layer. Figure 1 - Proposed AI-assisted workflow from image acquisition to geologist review and 3D model update. 4.3. Training and augmentation A first pilot cycle is defined with 500 images before augmentation, corresponding to 100 images for each of the five domain classes. This number is not presented as a final industrial dataset, but as a realistic starting point for a supervised pilot. Transfer learning from drill-core or hand-sample imagery can reduce the annotation burden because early CNN layers learn general visual primitives such as edges, colour gradients and textures. However, mine-face images require additional tuning due to lighting, wetness, support mesh and perspective distortion. Data augmentation should reproduce the actual underground scene. Recommended transformations include random crops, rotations, brightness variation, contrast variation, shadow simulation, blur, partial occlusion and wet-surface glare simulation. Augmentation is not used to artificially inflate confidence; it is used to expose the model to operational variability and to reduce overfitting to a limited set of training images. 4.4. Evaluation metrics Segmentation quality is evaluated with intersection-over-union (IoU) and Dice coefficient. For a predicted domain P and a geologist-validated domain G, IoU measures overlap relative to the union of both areas, while Dice emphasizes agreement relative to the combined size of the predicted and reference regions. IoU = |P intersection G| / |P union G| (1) Dice = 2 |P intersection G| / (|P| + |G|) (2) Object detection is evaluated with mean average precision at an IoU threshold of 0.5 (mAP@0.5). Operational performance is evaluated separately through review acceptance rate, mapping-cycle time and
RkJQdWJsaXNoZXIy MTM0Mzk2