Maritime Visual Classification at Smartports
Deep learning for maritime port operations — multi-task classification (ship detection, docking status) using PyTorch with ResNet18, exhaustive ablation studies, and real-time inference optimisation for smart port environments.
Visit websiteContext
Smart ports deploy camera networks to monitor vessel traffic, berth occupancy, and cargo operations. Manual monitoring is impractical at scale. This project builds a deep learning system that simultaneously classifies ship type and docking status from a single camera feed, enabling automated port management decisions in real time.
Model Design
- Backbone: ResNet18 pre-trained on ImageNet, fine-tuned with a lower learning rate on the first layers and full rate on the classification heads
- Multi-task output: Two parallel classification heads sharing the same feature extractor — one for vessel type, one for berth occupancy — trained with a weighted sum of cross-entropy losses
- Ablation study: Systematic comparison of frozen vs. fine-tuned backbone, single-task vs. multi-task training, and data-augmentation strategies
- Inference optimisation: TorchScript export and batch normalisation folding for low-latency deployment on edge hardware
Ablation Results (Accuracy)
| Configuration | Ship Type | Docking Status |
|---|---|---|
| Frozen backbone, single-task | 81.2% | 78.4% |
| Fine-tuned, single-task | 88.7% | 85.1% |
| Fine-tuned, multi-task (final) | 91.3% | 89.6% |