<![CDATA[Crack detection in photogrammetric mining imagery remains challenging due to severe class imbalance, thin and irregular crack morphology, and complex soil textures that hinder accurate boundary delineation. This study proposes a DeepLabV3+ semantic segmentation model with a ResNet-50 backbone trained using a Composite Loss that integrates Focal Tversky Loss, Weighted Cross-Entropy Loss, and Boundary Loss into a unified multi-objective framework. The dataset comprised 425 photogrammetric crack images acquired from open-pit mining sites, divided into 380 training and 45 test images. Data augmentation, including horizontal flip, vertical flip, channel shuffle, and random brightness and contrast adjustment, was applied to improve model generalization under varying image conditions. The proposed configuration was evaluated against Default Loss, Binary Cross-Entropy Loss, and Weighted Cross-Entropy Loss. The model achieved a Global Accuracy of 0.9808, Mean Intersection over Union of 0.7890, and Mean Boundary F1-Score of 0.8642, outperforming all baseline configurations. Boundary fidelity showed the largest improvement, with a Mean BFScore gain of 27.3% over Default Loss, demonstrating the effectiveness of Boundary Loss for precise crack contour delineation. The trained model was also deployed in a GUI-based system supporting automated crack analysis and PDF report generation for practical field application.]]>
Copyrights © 2026
