Underwater Image Segmentation (SUIM Dataset)

Semantic segmentation of underwater imagery using U-Net + EfficientNet-B0 encoder.

Dataset

SUIM Dataset on Kaggle

Classes (8 total)

Background, Human divers, Aquatic plants, Wrecks/Ruins, Robots/instruments, Reefs/invertebrates, Fish, Sea-floor

Model

• Architecture: U-Net (via segmentation_models_pytorch)
• Encoder: EfficientNet-B0 (ImageNet pretrained)
• Loss: Dice + Focal (combined)
• Input size: 128×128

Results

Best Validation Dice Score: 0.2662

How to Run

Open Underwater_imagery.ipynb in Google Colab and run all cells.