Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/tanim-mishkat/improved-image-classification-using-mobilenetv2

Butterfly Image Classification improved model with MobileNetV2
https://github.com/tanim-mishkat/improved-image-classification-using-mobilenetv2

colab-butterflies computer-vision data-augmentation deep-learning image-classification keras mixed-precision mobilenetv2 tensorflow transfer-learning

Last synced: 3 months ago
JSON representation

Butterfly Image Classification improved model with MobileNetV2

Awesome Lists containing this project

README 

          
# 🦋 Assignment Project: Butterfly Image Classification with MobileNetV2



This repository documents the implementation of a **deep learning pipeline for butterfly species classification**, completed as part of an AI course assignment. The assignment required the following workflow:



1. **Choose a Topic** → I selected *Butterfly Image Classification* within computer vision.

2. **Review Literature** → Read and summarized **10 research papers** related to image classification, transfer learning, and accuracy improvement techniques.

3. **Dataset Selection** → Identified a butterfly species dataset from Kaggle.

4. **Baseline Code** → Chose an existing Kaggle notebook implementation.

5. **Reimplementation & Improvement** → Rewrote the code into a clean, modular pipeline with enhancements in data quality, augmentation, evaluation, and accuracy.



---



## 📌 Project Overview



The goal of this project is to classify butterfly images into multiple species using a **transfer learning** approach with **MobileNetV2**. The workflow follows two-stage training:



* **Stage 1:** Train a new classification head while keeping the MobileNetV2 backbone frozen.

* **Stage 2:** Unfreeze the top layers of the backbone and fine-tune with a reduced learning rate.



This staged approach allows the model to preserve pretrained ImageNet features while adapting to butterfly-specific patterns.



---



## ✨ Key Features



* **Dataset Handling**: Supports both folder-structured and CSV-structured datasets. Includes checks for corruption and duplicates (SHA-1 hashes). Class imbalance addressed with computed class weights.

* **Preprocessing & Augmentation**: Images resized to 224×224, normalized to \[0,1], cached for faster training. Augmentations include random flip, rotation, zoom, contrast, and translation.

* **Transfer Learning**: MobileNetV2 backbone pretrained on ImageNet with custom classification head: `GlobalAveragePooling2D → Dropout → Dense(softmax)`.

* **Training Strategy**: Two-stage pipeline with OOM-safe batch fallback, AdamW optimizer, label smoothing, and learning rate schedules.

* **Regularization**: Dropout, L2 weight decay, and early stopping ensure generalization.

* **Evaluation Tools**: Classification report, confusion matrix, hardest example mining, and automated metrics logging.

* **Performance Optimizations**: Mixed precision training, Colab GPU-ready, and reproducible seeds.



---



## 📊 Alignment with 8 Key Factors for ML/DL Accuracy



The improvements were guided by the **8 Key Factors for ML/DL Accuracy** provided in the assignment:



| Factor                                 | Implementation in Project                                                                                                  | Notes                                                       |

| -------------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------- |

| **1. Data Quality & Quantity**         | ✔ Removed duplicates and corrupt files
✔ Class weights for imbalance
✔ Consistent preprocessing (resize + normalize) | ✘ Limited to Kaggle dataset; no extra data collected        |

| **2. Feature Engineering**             | ✔ Automated normalization and scaling
✔ Dimensionality reduction via GAP                                                | ✘ No manual feature creation (CNN learns features directly) |

| **3. Model Selection & Architecture**  | ✔ MobileNetV2 backbone
✔ Shallow, regularized head design                                                               | ✘ Did not compare with other architectures due to time      |

| **4. Hyperparameter Tuning**           | ✔ AdamW optimizer
✔ Label smoothing
✔ Batch size fallback
✔ ReduceLROnPlateau                                     | ✘ No full hyperparameter sweep (compute-limited)            |

| **5. Data Augmentation**               | ✔ Applied augmentations (flip, rotation, zoom, etc.)
✔ Implemented MixUp (optional)                                     | ✘ Text/audio augmentation not applicable                    |

| **6. Regularization & Generalization** | ✔ Dropout (0.25)
✔ Weight Decay
✔ Early Stopping                                                                     | ✘ Ensemble methods not implemented                          |

| **7. Evaluation & Feedback Loop**      | ✔ Stratified validation split
✔ Confusion matrix & hardest examples
✔ JSON/CSV/PNG artifacts saved                   | ✘ No automated retraining loop                              |

| **8. Computational Resources**         | ✔ Mixed precision for GPU efficiency
✔ Early stopping for time savings                                                  | ✘ Did not use TPUs or multi-GPU training                    |



---



## 🚀 Getting Started



### Run in Google Colab



1. Upload your `kaggle.json` API key.

2. Run notebook cells sequentially.

3. Dataset auto-downloads and extracts.



### Local Setup



```bash

# Clone repository

git clone https://github.com/your-username/butterfly-image-classification-mobilenetv2.git

cd butterfly-image-classification-mobilenetv2



# Create environment

python -m venv .venv

source .venv/bin/activate   # Windows: .venv\Scripts\activate



# Install dependencies

pip install -r requirements.txt

```



---



## 🧠 Model Architecture



* **Backbone:** MobileNetV2 (ImageNet pretrained)

* **Head:** GlobalAveragePooling2D → Dropout(0.25) → Dense(NUM\_CLASSES, softmax with L2 regularization)

* **Optimizer:** AdamW (lr=3e-4, wd=1e-4)

* **Loss Function:** Categorical Crossentropy (label smoothing=0.1)

* **Training:** Two stages (frozen backbone → fine-tuning top \~40%)



---



## 📈 Training Process



1. **Stage 1:** Train only the classification head with backbone frozen.

2. **Stage 2:** Unfreeze last 40% of MobileNetV2 layers, reduce learning rate, fine-tune.

3. **Reason for Two Stages:** Directly training the whole model can destroy pretrained features; staged training preserves general patterns while adapting to butterfly images.



---



## 📊 Evaluation Artifacts



All evaluation results are stored under `results/`:



* `metrics.json` — class-level + overall accuracy

* `confusion_matrix.png` — visual class confusions

* `hardest_examples_topN.json` — misclassified high-loss examples

* `test_predictions.csv` — predictions for unlabeled test set



---



## 🛠 Requirements



* Python 3.9+

* TensorFlow 2.15+

* scikit-learn

* pandas, numpy, matplotlib

* TensorFlow Probability (optional, for MixUp)



---



## 🗺 Future Improvements



* Add support for **EfficientNet** and other backbones

* Perform **hyperparameter sweeps** for LR, dropout, smoothing

* Enable **test-time augmentation (TTA)**

* Explore **ensembles** for improved robustness



---



## 📝 License



MIT License