Rock fragment boundary detection using compressed random features

Sections of the mining industry depend on regular analysis of rock fragmentation to detect trends that may affect safety or production. The limitations inherent in 2D imaging analysis mean that human input is typically needed for delineating individual rock fragments. Although recent advances in 3D image processing have diminished the need for human input, it is often infeasible for many mines to upgrade their existing 2D imaging systems to 3D. Hence there is still a need to improve delineation in 2D images. This paper proposes a method for delineating rock fragments by classifying compressed Haar-like features extracted from small image patches. The optimum size of the image patches and the number of compressed features are determined empirically. Experimental results show the proposed method gives superior results to the commonly used watershed algorithm, and compressing features improves computational efficiency such that a machine learning approach is practical.