In the process of iron ore agglomeration, there are strict requirements for the pelletizing particle size. In order to detect the pelletizing particle size in real time, based on the analysis of the dynamic region image of the balling device, a pelletizing segmentation algorithm on foreground extraction and adhesion segmentation is proposed. In order to solve the problem of mineral powder interference in the background of image pelletizing, on the basis of Unet, a separable convolutional structure is integrated, and a lightweight DSC-Unet foreground segmentation network is designed to ensure the robustness of foreground segmentation. At the same time, in view of the overlapping problem of pelletizing, a watershed segmentation method based on distance transformation is used to realize the segmentation of foreground adhesion, and then the diameter of the pelletizing is obtained by circular fitting. The experimental results show that the segmentation accuracy of the algorithm involved reaches 97%, which is better than other existing algorithms and meets the detection requirements.