A Lightweight Dual-Branch Network with Adaptive Feature Fusion for Particle Segmentation in the Pelletization Process

Particle size distribution (PSD) is a critical parameter in industrial fields such as fertilizer production and mineral processing, directly impacting product quality and production efficiency. While deep learning techniques have advanced PSD analysis, they often face limitations in addressing the specific demands of particle analysis, particularly in achieving a balanced multi-scale representation, enabling effective cross-layer feature integration, and maintaining computational efficiency. To address these challenges, this work proposes a lightweight dual-branch encoder with adaptive weight feature fusion, aiming to achieve high-accuracy and efficient particle image segmentation. The encoder combines a lightweight CNN with a MobileNetV3 network, improving multi-scale feature extraction capabilities. Furthermore, an adaptive fusion mechanism is incorporated during the decoding stage to optimize detail preservation. Experimental results on particle segmentation in the pelletization process demonstrate that the proposed method surpasses existing approaches in both segmentation accuracy and computational efficiency, offering a highly efficient and reliable solution for real-time industrial PSD monitoring.