Nigel C. Shillingford, Ana Rochelle L. Quinal, Christian V. Maderazo
Microplastic pollution poses growing environmental and health risks. Existing approaches often rely on laborious laboratory techniques or computationally intensive models that require high-end hardware. This paper proposes a resource-efficient dual-branch detection model, combining a Swin Transformer branch for global context extraction with a convolutional neural network branch for local feature representation, applied to microscopic images of microplastics. The detection model was trained on a publicly available augmented dataset and achieved a mean average precision of 90.01% and an F1-score of 0.9492, while using only 0.54 million parameters-Approximately five to six times fewer than comparable YOLO-based models. The architecture operates at 214.6 FPS with a VRAM footprint of 24.71 MB, enabling deployment on consumer-grade hardware. Overall, the detection model achieves a strong trade-off between efficiency and detection performance, making it viable for deployment on resource-constrained devices. © 2026 IEEE.
University of San Carlos, Department of Computers Information Sciences and Mathematics, Cebu City, Philippines
