The pellet size is one of the main factors affecting blast furnace air permeability，blast furnace smelting efficiency and energy consumption. In order to solve the problem that the pellet size is difficult to measure accurately under industrial conditions，the Mask R-CNN model combined with attention mechanism is used to segment and measure the pellet size. After preprocessing the pellets image，the pellet dataset is constructed，the training performance of a variety of backbone networks is compared，and the accuracy is compared with multiple segmentation models. In addition，the pellet size is measured by using the pixel statistically segment mask area. The results show that ResNet50 is superior as the backbone network in feature extraction of pellets. The Mask R-CNN model of Convolutional Block Attention Module ( CBAM) is introduced，which improve by 2. 18% compared to the initial model A_mean . Compared with BlendMask，SOLOv2，YOLAT and CondInst，the improved model also has advantages in segmentation accuracy and can better handle segmentation details. In addition，compared with the pellet size measured by Image J，the maximum error of the pellet size measurement method proposed in this paper is kept within ± 1. 8 mm，and A_{IoU = 0. 5} can reach 0. 948 3.