Microplastic pollution poses escalating risks to aquatic ecosystems and human health, yet existing analytical methods such as microscopy and spectroscopy remain too slow and costly for large-scale surveillance. We introduce FasterRes-FPN, an end-to-end detector that couples Faster R-CNN with a ResNet-50 feature-pyramid backbone. Trained on the publicly released microplastic_100 dataset (320 train/80 val images in COCO format), the model simultaneously classifies and localises four common particle types—fiber, fragment, pellet, and film—under challenging conditions of turbidity, lighting variation, and background clutter. On the held-out split FasterRes-FPN achieves an F1 score of 0.912, \(m\!AP_{0.5}=0.901\) , and recall 0.948, outperforming strong one-stage YOLO baselines in recall while maintaining competitive precision. Error analysis pinpoints thin films as the primary failure mode, driven by low contrast and amorphous boundaries. These results demonstrate that two-stage deep detectors can deliver near-real-time, expert-free assessment of microplastic abundance in natural waters. Future work will target improved localisation of sub-millimetre films, domain adaptation to deep-sea and wastewater imagery, and compression for deployment on autonomous surface and underwater platforms.