Abstract
After blasting in an open-pit mine, it has great guiding significance for the subsequent secondary crushing, shovel loading, transportation and other processes to obtain the large ore fragmentations of the blasting pile, which also plays an important role in improving the efficiency and economic benefits of the mine. In this paper, a large ore detection and measurement model LODM based on Mask R-CNN is proposed. After training on our MPBRD1.0 dataset, we compare the detection results with traditional image segmentation algorithms: the K-means clustering algorithm, Canny edge detection algorithm, watershed algorithm and ore image segmentation algorithm based on the U-Net network, which proves that the detection results of the LODM model are more in line with the actual situation. To improve the detection ability of the LODM model, we propose a ResNet34 feature extraction network as the backbone and train ResNet50, ResNet101 and VGG16 at the same time. The results show that the performance of the LODM model can be optimized by using the ResNet34 feature extraction network.