A non-adaptive segmentation algorithm for particle images in controlled environments with uniform backgrounds based on two-round superpixel segmentation and ensemble learning

Abstract

Particle image segmentation under controlled environments with uniform backgrounds remains a challenging task due to issues such as particle adhesion, low contrast, and uneven illumination. Existing methods often suffer from over-segmentation or under-segmentation, especially when applied to microscopic or industrial particles. To address these problems, this paper proposes a non-adaptive segmentation algorithm called TS-EL (Two-round Superpixel Segmentation and Ensemble Learning), which is specifically designed for particle images captured in controlled settings with homogeneous backgrounds. The TS-EL framework performs coarse-to-fine superpixel segmentation and easy-to-hard classification. It introduces a gradient distance-based superpixel segmentation algorithm (GradSE) to improve boundary alignment between superpixels and particle contours. A Gaussian model and dual-factor classification criteria are employed to categorize high-confidence superpixels into foreground and background, while low-confidence regions are refined using a second-round segmentation based on minimum bounding boxes. The final classification of ambiguous regions is achieved via the LogitBoost ensemble learning algorithm. Experimental results on three types of particle images (grain, color masterbatch, and cell images) demonstrate that the proposed method outperforms seven state-of-the-art comparative algorithms in terms of segmentation accuracy and boundary adherence. The method is non-adaptive and relies on empirically set parameters, making it well-suited for batch processing in controlled environments but less generalizable to natural or complex scenes.