Abstract Iron ore pellets are one of the main raw materials in the metallurgical industry. To ensure the utilization rate of raw materials and the efficiency of industrial processes, it is often necessary to control the particle size distribution of the pellets. Computer vision-based methods for detecting iron ore pellet size distribution have been developed. However, pellet shadowing, particle overlap, and uneven illumination pose significant challenges that severely compromise segmentation performance and particle size measurement accuracy. To address these challenges, we proposed an improved network structure called Transformer-Fused Level Set UNet3+ (TFL-UNet3+). Specifically, UNet3+ provides multi-scale feature representations to automatically initialize the level set curve, while a reconstructed energy functional, substituting traditional pixel intensity terms with the discrepancy between the probability map and the ground truth mask, is embedded into the loss function for joint supervision of semantic and geometric constraints. Moreover, the level set evolution is reformulated as a temporal sequence, enabling the Transformer to capture long-range dependencies across iterations and thus alleviating degradation and instability of level set. Experimental evaluations on disc pelletizer discharge images demonstrate that the proposed TFL-UNet3+ achieves superior segmentation accuracy, yielding relative improvements of 1.3(%) in Intersection over Union (IoU) and 0.9(%) in Boundary F1 Score (BF Score) over state-of-the-art methods, while maintaining reliable PSD measurements validated.
Deng et al. (Tue,) studied this question.