README.md

metadata

title: Fish Freshness Classifier
emoji: 🐟
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 4.19.2
app_file: fish_freshness_app.py
pinned: false

🐟 Food Freshness Classification Using Deep Learning

This project aims to classify food images as Fresh or Not Fresh using deep learning and transfer learning with PyTorch. We explore various CNN architectures and hybrid ensemble models to achieve state-of-the-art accuracy.

---

📁 Project Structure

├── models/ │ ├── efficientnet_b0.py │ ├── efficientnet_b4.py │ ├── mobilenetv2.py │ ├── resnet50.py │ └── hybrid_fusion.py │ ├── train/ │ ├── train_efficientnet.py │ ├── train_resnet50.py │ ├── train_mobilenetv2.py │ ├── train_efficientnet_b4.py │ └── train_hybrid_fusion.py │ ├── evaluate/ │ ├── evaluate_efficientnet.py │ ├── evaluate_resnet50.py │ ├── evaluate_mobilenetv2.py │ ├── evaluate_efficientnet_b4.py │ ├── evaluate_hybrid_fusion.py │ └── compare_fish_models.py │ ├── utils/ │ └── dataset_loader.py │ ├── data/ │ ├── train_paths.npy │ ├── train_labels.npy │ ├── val_paths.npy │ ├── val_labels.npy │ ├── test_paths.npy │ └── test_labels.npy │ ├── results/ │ ├── eval_metrics_.png │ ├── confusion_matrix_.png │ ├── model_metrics_summary.csv │ └── *.pth (saved models) │ ├── requirements.txt └── README.md

---

🧠 Model Architectures

✅ Individual Models:

• EfficientNetB0
• EfficientNetB4
• MobileNetV2
• ResNet50

Each uses transfer learning:

• Initial training with train_base=False (frozen feature extractor)
• Followed by fine-tuning the deeper layers

🔁 Hybrid Model:

EnhancedHybridFusionClassifier

• Combines EfficientNetB0, ResNet50, and MobileNetV2
• Fuses feature embeddings via concatenation
• Classifier head with multiple dense layers
• Fine-tuned after head training

---

🧪 Evaluation and Comparison

Each model is evaluated on the same test set.

Metrics:

• Accuracy
• Precision
• Recall
• F1 Score
• AUC (ROC)

Visual Outputs:

• confusion_matrix_*.png
• eval_metrics_*.png
• ROC curves and metric bar charts (compare_fish_models.py)

Results are also exported to: results/model_metrics_summary.csv

---

🏗️ Training Details

• Optimizer: Adam
• Loss: BCELoss
• Scheduler: ReduceLROnPlateau (monitors val accuracy)
• Early stopping via logic inside training script
• Data augmentation during training:
  • Random horizontal flip
  • Rotation
  • Color jitter
• num_workers optimized based on CPU
• pin_memory=True for faster GPU transfer

---

💾 Setup & Requirements

1. Clone the repository:

git clone <your-repo-url>
cd <repository-name>

2. Create and activate virtual environment:

# Windows
setup.bat

# Or manually:
python -m venv venv
venv\Scripts\activate  # Windows
source venv/bin/activate  # Linux/Mac
pip install -r requirements.txt

Requirements:

• Python 3.8+
• PyTorch 2.0+
• CUDA-compatible GPU (recommended)
• See requirements.txt for full list

🚀 How to Run

1. Train a Model:

python train/train_efficientnet.py

2. Evaluate:

python evaluate/evaluate_efficientnet.py

3. Compare All Models:

python evaluate/compare_fish_models.py

🏁 Results Summary

Model	Accuracy	Precision	Recall	F1
EfficientNetB0	0.9819	0.9652	0.9978	0.9812
EfficientNetB4	0.9797	0.9733	0.9843	0.9788
MobileNetV2	0.9562	0.9410	0.9685	0.9546
ResNet50	0.9765	0.9648	0.9865	0.9756
HybridFusion	0.9755	0.9627	0.9865	0.9745

📌 Notes

• Training logs are printed per epoch with loss/val accuracy
• Best models are saved as *_best.pth
• Hybrid model training uses transfer learning and feature fusion
• Comparison across architectures and training strategy is fair (same data, same pipeline)

🏆 Final Thoughts

This project shows the power of transfer learning and hybrid deep learning models in food quality assessment. The modular pipeline supports extension (e.g., Grad-CAM, more ensembling) and can serve as a template for similar classification tasks.

Acknowledgments

• Segment Anything Model (SAM)
• EfficientNet Implementation