A pellet segmentation and particle size measurement model based on improved YOLOv11

In the blast furnace smelting process, the uniformity of pellet size has a significant impact on the smooth operation of the blast furnace. However, in the industrial pelletizing process, traditional manual screening methods are still used on-site. These methods are inconvenient for direct and continuous measurement of pellet size and have large measurement errors, making it difficult to meet the requirements of real-time measurement. Therefore, this paper proposes a non-contact online method for measuring pellet size. Based on the YOLOv11 model, a multi-scale enhanced upsampling module (MEUM) is introduced to replace the original upsampling structure. Multi-scale feature fusion and edge enhancement design improve the representation ability for different particle sizes and fuzzy boundaries. A local importance attention (LIA) mechanism is embedded in the backbone network to adaptively enhance the response of key regions, improving the robustness of target features in complex backgrounds while maintaining the network's lightweight nature. Furthermore, on this basis, an edge contour-Hough circle joint detection method is proposed to calculate the pellet size. The minimum circumscribed circle is obtained by extracting the edge contour of the processed mask, and then matched and screened through multi-scale Hough circle detection to obtain a stable and reliable fitted circle and calculate the particle size. The results show that the mAP_50-95(The average of the average accuracies calculated under multiple thresholds ranging from 0.50 to 0.95 for IoU) of box detection of the improved model increases from 0.885 to 0.906. The recall rate of mask detection increases from 0.993 0 to 0.999 7, and its mAP_50-95 increases from 0.833 to 0.847. Compared with the other four comparative models, this model achieves the optimal performance. Meanwhile, the maximum error between the proposed pellet size measurement method and the ImageJ measurement method remains within ±1.7 mm, with an average relative error of 3.98%. The proposed pellet size detection method can efficiently handle pellet particle size identification tasks in complex industrial environments and has broad application prospects in the field of intelligent metallurgical industry, providing new ideas and methods for non-contact pellet size detection.