Investigation of Object Detection Based Method for Open-Pit Blast Quality Estimation

This paper deals with the investigation of open-pit blast quality estimation methods using computer vision systems. The methods are based on the estimation of the size (or area) distribution of the obtained rock fragments. Analysis of proposed literature methods based on instance segmentation was carried out. We demonstrate that the results of the object detection-based methods correspond to those estimated by the instance segmentation method with regard to the maximum size of the rock fragments. This result was approved statistically. In this paper it is shown that object detection-based methods allow to speed up the estimation task in tens of time. This allows reaching a trade-off between accuracy and computational complexity, which is the up-to-day task for poorly performing (low-price) devices. This paper shows how the YOLOv5-based models can reach a comparable level of accuracy as the traditional Mask-RCNN-based models with an inference time approximately 10 times shorter (about 1.5 s for the poorly performing CPU-device).