Leaf Disease Segmentation

Binary segmentation of diseased regions on plant leaves using a UNet architecture with a pretrained MobileNetV2 encoder. Trained on the PlantSegV2 dataset.

Results

Columns: Input Image | Ground Truth Mask | Predicted Mask

The model segments diseased regions well on most samples. Some over-prediction occurs on complex backgrounds, likely due to similar texture/color between disease and background.

Architecture

• Model: UNet + MobileNetV2 encoder (ImageNet pretrained)
• Loss: BCE + Dice (50/50)
• Input size: 256×256
• Classes: 1 (binary — disease vs background)

Dataset

• Source: PlantSegV2 (Kaggle)
• Size: ~11k images

Setup

pip install -r requirements.txt

Train

python3 train.py                           # default config.yaml
python3 train.py --config my_config.yaml  # custom config

All hyperparameters are in config.yaml.

Evaluate

python3 eval.py --split test
python3 eval.py --split test --n 10 --save-dir ./eval_output

Inference basic UI

python3 app.py

Launches a Gradio web interface. Upload a leaf image and get the overlay and the predicted mask. Must have a trained model under: ./checkpoints/best.pth