This project focuses on classifying leaf images to predict plant species using Deep Neural Networks. We experimented with various techniques, including custom networks and transfer learning with fine-tuning, to achieve high accuracy in leaf classification.
- 17,728 images of leaves on a black background
- 14 plant species (apple, blueberry, cherry, corn, grape, orange, peach, pepper, potato, raspberry, soybean, squash, strawberry, tomato)
- 256x256 resolution JPEG images
- Unbalanced class distribution
- Data analysis and preparation
- Data augmentation techniques
- Transfer learning with pre-trained models
- Custom CNN architectures
- Performance comparison of various models
- Dataset splitting (training, validation, test)
- Oversampling and class weighting to address class imbalance
- Data augmentation (rotation, zoom, flipping, translations, shears)
- Transfer learning: VGG16, ResNet152v2, EfficientNetV7, DenseNet-201, Xception, ResNet50v2
- Custom deep convolutional neural networks
- Experiments with autoencoder-based feature extraction
- Cross-validation via hold-out
- Two-step training for transfer learning models
- Learning rate adjustment
Best performing models on remote test data:
| Model | Accuracy |
|---|---|
| DenseNet-201 | 0.9075 |
| EfficientNetV7 | 0.9038 |
| ResNet152v2 | 0.8736 |
| ResNet50V2 | 0.8264 |
| Xception | 0.7264 |
Transfer learning models with fine-tuning achieved the best performance in terms of accuracy and training time, given the limited training resources. DenseNet-201 showed the highest accuracy of 90.75% on the test set.
- TensorFlow
- Keras
- Scikit-Learn
- Jupyter Notebook
- Marco Domenico Buttiglione
- Luca De Martini
- Giulia Forasassi
Politecnico di Milano
November 30, 2021
For more detailed information about the methodology, experiments, and findings, please refer to the full project report.