Smartphone-Based Hybrid Deep Learning System for Real-Time Microplastic Detection in Drinking Water

Microplastic-contaminated drinking water poses a threat to the environment and human health, and laboratory detection approaches are still expensive, slow, and impractical even for daily use. In this article, we present a smartphone- based AI system for microplastic detection in real-time with lightweight deep-learning models and preprocessing. Image quality is improved with the use of a Super-Resolution module constructed using a shallow and lightweight upsampling network consisting of bilinear interpolation (nn. Up sample, scale_factor=2), and pixel-value clamping for efficient 2× image enhancement used to enable mobile deployment. Three models — YOLOv8, Faster R-CNN, and U-Net — had been tested, and the YOLOv8n was the one with the best performance (Precision ≈ 0.87, Recall ≈ 0.75, mAP@50 ≈ 0.84, mAP@50–95 ≈ 0.51), contributing to its chosen detection model. For on-device Android inference, the trained YOLOv8n model was exported via the Py-Torch Mobile Lite compiler by converting to Torch Script and optimizing it with optimize_for_mobile() to create a .ptl file. The end product isfast, low-power, cost-effective, and issuitable in the context ofscalable citizen-led environmental monitoring.