TEA-Watershed:A temporal-enhanced adaptive watershed framework for real-time particle size measurement in dynamic industrial flows

Abstract

Accurate real-time particle-size measurement in rapid, illumination-varying industrial flows is hindered by motion blur, inter-particle adhesion, and uneven grayscale distributions. This study introduces TEA-Watershed (Temporal-Enhanced Adaptive Watershed), a training-free framework that delivers robust in-line metrology without interrupting production. The algorithm fuses consecutive frames to reinforce edges, integrates motion-aware parameter optimization with adaptive Otsu thresholding inside a watershed pipeline, and iteratively refines segmentation through trajectory-based feedback. An unsupervised K-Means module further groups particles by morphology, eliminating manual annotation while maintaining calibration under changing flow conditions. Validation on 2–15 mm coal and iron-ore streams achieved a mean Intersection-over-Union of 89.1 % and pixel accuracy of 96.2 %. For metrological performance, the system recorded a mean absolute size error of 3.8 % and a repeatability coefficient of variation below 2.5 % across 100 replicates. Processing throughput reached 22 frames s⁻¹(45 ms per frame), enabling continuous monitoring. Correlation with ISO 2591 sieve analysis confirmed high reliability (R² = 0.98). Requiring only modest computational resources and no learned weights, TEA-Watershed provides quantified uncertainty within industrial tolerances, offering a practical, scalable solution for particle-size measurement, screening-efficiency assessment, and flow diagnostics in mineral processing, coal beneficiation, and powder-handling operations.