Abstract

Mining and construction environments are specialized sectors that involve complex operations, heavy machinery, and specific safety considerations. Properly managing different resources (e.g., minerals. concrete, and stones) is essential for successfully executing projects. In this regard, Deep Learning (DL) algorithms emerge as a catalyst in classifying and detecting several resources across the supply chain within hazardous mining and construction environments. This work presents a comparative analysis of different DL techniques used for detecting rocks in challenging mining environments. To this end, three models - MarsNet, YOLOv8+SAM, and SegFormer - have been designed, training and experimentally tested. In addition, we analyze the generalization of proposed algorithm according to selected dataset. Experimental findings disclose that the Segformer algorithm outperforms semantic segmentation algorithms based on other deep learning algorithms regarding main metrics such as recall, F1-score, and Intersection over Union.