Segmenting images of coal piles on a belt is an unsolved problem in coal-based machine vision research, though it is an essential step for estimating size distribution and classifying coal. In this investigation, a new algorithm for segmenting images of coal piles on a belt is proposed. A multi-scale linear filter, constructed of a Hessian matrix and Gaussian function, forms the core of this algorithm and obtains an edge intensity image to form good seed regions for a watershed segmentation. Manual segmentation is used to define ground truth segmentation images to quantify the errors of the proposed method. Tests indicate that 12.76% of the visible regions are under- or over-segmented, and that this algorithm is feasible and effective in practical applications.