Abstract: The blast furnace tuyere raceway is the dynamic core of ironmaking process, and real time changes in coke particle size directly reflect the thermal strength degradation of coke and the activity of hearth. However, traditional monitoring methods relying on furnace stoppage sampling have disadvantages such as strong sampling intermittence, large human random errors, and significant disruption to normal production, making it difficult to meet the intelligent and continuous monitoring demands of modern blast furnaces. To address this issue, this paper developed an online monitoring system for coke particle size at the blast furnace tuyere, aiming to achieve real time and accurate perception of coke geometric dimensions in closed high temperature environment using computer vision technology. The study proposed a lightweight detection model named ESG-YOLO （EffectiveSE GhostConv-YOLO）. This model introduced a GhostConv module in the backbone network for lightweight feature extraction, utilized an EffectiveSE mechanism to enhance edge perception, designed a weighted feature fusion （WFF） module to capture spatial correlations of cross scale features, and adopted the Focal-PIoU2 loss function to optimize localization regression. Experimental results show that the improved ESG-YOLO model reduces computational cost by 33.33% while increasing the mean average precision P_mA50 and P_mA50：95 to 98.10% and 90.90% respectively, and raising the F₁ score to 96.40%. These improvements significantly enhancing the model's robustness in low signal to noise ratio environments with high temperature, high brightness, and pulverized coal injection interference. Industrial field validation demonstrates that the system achieves an average inference speed of 106.8 F/s, enabling real time capture of coke particle size evolution during unsteady processes such as furnace stoppage and restart. By analyzing the dynamic correlation between circumferential particle size standard deviation and the permeability index, the result indicates that particle size fluctuation can serve as a quantitative prior indicator for characterizing circumferential inhomogeneity in the hearth. The application of this system effectively reduces manual sampling intensity and provides scientific basis and engineering support for online evaluation of tuyere raceway activity using machine vision.