The particle size of ore is an important reference to judge the crushing effect of crusher, and image segmentation is the key step of ore particle size detection. To solve the problems of image segmentation inaccuracies caused by complex shape, adhesion and stacking of broken ore, and serious image noise, a broken ore image segmentation method based on improved HED (Holistically-Nested Edge Detection) network model is proposed. First, the bilateral filtering pre-processing operation is carried out on the collected ore image to reduce the influence of noise on segmentation. Second, the residual deformable convolution block is used to replace the ordinary convolution block to enhance the feature extraction ability of the model for ores of different sizes and shapes, and the void convolution is used to replace the original pooling layer to expand the receptive field and retain the global information of ores. Finally, the HED network framework with a bottom-short connection structure is used for feature extraction of ore, and the extracted features are combined with low-level detail information to reduce the problem of undersegmentation of cohesive and stacked ore particles.