中国矿山工程

In blasting engineering, the blasting lumps size are the key index to evaluate blasting effect and optimize blasting parameters. The key to the blasting lumps size recognition is to fully extract the effective feature information of complex rock image, and then accurately segment the image. The serious stacking between ore and rock is the main reason for the inability of accurate segmentation. The traditional image segmentation algorithms are usually based on threshold, edge detection and so on. These methods have the defects of poor accuracy, weak robustness and low practicability. Combined with the actual needs of mine blasting pile block recognition, an improved mask RCNN ore rock segmentation model is proposed. The noise reduction image of the protected edge is generated by bilateral filtering, and the noise reduction image is used as the data set of mask RCNN to generate a learning model for ore rock segmentation. Compared with the traditional OTSU and Canny edge detection algorithms, this algorithm has high accuracy, strong practicability and better robustness; Compared with UNET algorithm, it solves the local segmentation defect of UNET on the premise of ensuring the segmentation accuracy. The experimental results show that the pixel difference curves of manual segmentation and model segmentation fit well, which has a certain engineering significance for the identification of the blasting lumps size recognition.