Herbify is a comprehensive deep learning project focused on the classification of 91 types of herbs using advanced computer vision techniques. The project includes data preprocessing, augmentation, model training (with various state-of-the-art architectures), hyperparameter optimization, and ensemble learning.
-
Augmentation/: Jupyter notebooks for image augmentation using Albumentations and OpenCV. Includes multiple strategies (A, B, C) for increasing dataset diversity.
-
Data_Split/: Notebook for splitting the dataset into train, test, and validation sets (default: 70/15/15 split).
-
PAHD/: Stands for Preprocessing Algorithm for Herb (object) Detection. Notebooks for advanced preprocessing, including background removal, contour extraction, and adaptive histogram equalization. Includes single-image and batch processing examples.
- Optimal_Hyperparameters_Grid_Search.ipynb: Grid search for optimal batch size, learning rate, and weight decay. Includes training, validation, and test evaluation for each configuration.
Contains subfolders for each model family, each with a notebook for transfer learning and fine-tuning on the herb dataset:
- EfficientNet/:
EfficientNet_v2_Large.ipynb - MobileNet/:
MobileNet_v3_Large.ipynb - ResNet/:
ResNet152.ipynb - VGG/:
VGG16.ipynb,VGG19.ipynb - ViT/: Vision Transformer models (
ViT_B_16_e2e.ipynb,ViT_B_16_v1.ipynb,ViT_L_16_e2e.ipynb,ViT_L_16_v1.ipynb)
Each notebook covers:
- Data loading and preprocessing
- Model setup and architecture (along with parameter freezing, early stopping and other configurations)
- Transfer learning (or Fine-tuning)
- Training, validation, and evaluation (accuracy, F1, precision, recall, confusion matrix)
Jupyter notebooks for combining predictions from multiple models to improve classification performance. Includes various ensemble strategies:
-
MobileNet v3 Large + ViT-B/16
-
VGG19 + ViT-B/16
-
ResNet152 + ViT-L/16
-
EfficientNet v2 Large + ViT-L/16
-
MobileNet v3 Large + VGG19 + ViT-B/16
-
ResNet152 + EfficientNet v2 Large + ViT-L/16
-
MobileNet v3 Large + VGG19 + EfficientNet v2 Large + ViT-L/16
-
VGG19 + ResNet152 + EfficientNet v2 Large + ViT-L/16
Multi-model ensembles (e.g., ResNet, EfficientNet, VGG19, ViT)
-
Requirements: Python 3.8+, PyTorch, torchvision, scikit-learn, albumentations, OpenCV, matplotlib, tqdm, colorama, PIL, numpy, pandas, scikit-image.
-
Data: Place your herb image dataset in the appropriate directory as referenced in the notebooks.
-
Run Notebooks: Follow the order: Preprocessing → Hyperparameter Search (for best models) → Fine-Tuning → Ensemble.
- Flexible Preprocessing: Advanced background removal and augmentation.
- Multiple Architectures: Fine-tuning of CNNs (VGG, ResNet, EfficientNet, MobileNet) and Transformers (ViT).
- Ensemble Learning: Combine strengths of different models.
- Reproducibility: Fixed random seeds and clear data splits.
- DIMPSAR: DIMPSAR Dataset
- DeepHerb: DeepHerb Dataset
- Herbify: Custom dataset of 91 herbs, available on request.
If you use this project, please cite appropriately.
This project is for academic and research purposes licensed under the MIT License. For commercial use, please contact the author.