Segmentation of mine overburden dump particles from images using Mask R CNN

Abstract

The stability of mine overburden dumps is crucial for the efficient operation of mining industries. The size distribution of particles affects the shear strength of dump slopes. Identification of dump particles from images is challenging as they vary in size, shape, color, granularity, and texture. In this paper, a unique way of identifying the particles from dump images using Artificial Intelligence is presented that can be used to determine the particle size distribution of dump. Mask R CNN with ResNet50 plus an FPN as a backbone network which is the current state of the art for instance segmentation has been implemented to segment the particles from dump images at detailed pixel level and to obtain their boundary. Experimental results showed promising results to delineate the particles and obtain masks over them. Our model has achieved a training accuracy of 97.2% for the dataset containing 31,505 particles. The model predicted the areas of dump particles with a mean percentage error of 0.39% and a standard deviation of 0.25 when compared to the ground truth values. The calculation of coordinates of the detected boundaries using the model significantly reduces the time and effort that are generally put in rock mechanics laboratories.