FreshCheck AI – Fresh vs Spoiled Fruit Classifier

A deep learning–powered application to classify fruits as fresh or spoiled.

Model: ResNet-50 (Transfer Learning)   |   Framework: PyTorch   |   UI: Streamlit

This project demonstrates an end-to-end computer vision workflow — from data preparation and leakage-safe splitting to model training, evaluation, and deployment using Streamlit.

---

A practical deep learning project that classifies fruit images as fresh or spoiled using a pretrained ResNet-50 model.
The project emphasizes data integrity, correct evaluation, and production-ready deployment, rather than just high accuracy.

---

📸 Screenshots

🖥️ Streamlit Application Interface

✅ Fresh vs Spoiled Result

---

✨ Features

• Binary Classification: Predicts whether a fruit is fresh or spoiled
• Multi-Fruit Support: Trained on 8 different fruits
• Transfer Learning: Uses pretrained ResNet-50 for efficient learning
• Leakage-Safe Dataset Splitting: Fruit-aware and class-aware splits
• Minimal Hyperparameter Tuning: Lightweight Optuna demonstration
• Fast Inference: Backbone frozen for quick predictions
• Interactive UI: User-friendly Streamlit web app

---

📂 Project Structure

Fruits_Freshness_Classifier/
│
├── README.md          
├── .gitignore
└──streamlit_app/
    │
    ├── artifacts/
    ├── harvest_classifier_notebook.ipynb
    └── harvest_classifier_resnet50.ipynb
    │
    ├── models/
       └── fruits_classifier_resnet50_tl.pth
    ├── app.py
    ├── app_v2.py # ✅ Main Streamlit application (use this)
    ├── model_definition.py # Model architecture definition
    ├── model_helper.py # Model loading & inference utilities
    ├── requirements.txt

---

🧠 Model Details

• Architecture: ResNet-50 (pretrained on ImageNet)
• Training Strategy:
  • Transfer learning
  • Backbone frozen
  • Fully connected layer fine-tuned
• Classes:
  • 0 → Fresh
  • 1 → Spoiled
• Input Size: 224 × 224 RGB
• Loss Function: CrossEntropyLoss
• Optimizer: Adam / AdamW (with weight decay)

---

🖥️ Setup Instructions

Prerequisites

• Python 3.9+
• Git

---

1️⃣ Clone the Repository

git clone https://github.com/inv-fourier-transform/Fruits_Freshness_Classifier.git
cd Fruits_Freshness_Classifier

2️⃣ Create & Activate Virtual Environment

python -m venv venv
# Windows
venv\Scripts\activate
# macOS / Linux
source venv/bin/activate

3️⃣ Install Dependencies

pip install -r requirements.txt

🚀 Running the Application

streamlit run app_v2.py

📸 How to Use the App

1. Upload a fruit image
2. The model processes the image
3. View the output, which includes:
   • Prediction: Fresh / Spoiled
   • Confidence Score

---

🔬 Training & Experiments

All experimentation and training logic is documented in the notebooks inside the artifacts/ directory.

Includes:

• From-scratch CNN experiments
• BatchNorm & Dropout analysis
• Data leakage diagnosis
• Transfer learning with ResNet-50
• Minimal Optuna-based tuning (demo only)

---

📦 Model Reuse

The trained model can be loaded independently:

import torch

model = torch.load("models/fruits_classifier_resnet50_tl.pth")

Refer to model_helper.py for reusable inference utilities.

---

🔐 Key Design Principles

• Correct evaluation > inflated accuracy
• Data leakage prevention
• Minimal but meaningful regularization
• Clear separation of training and inference
• Production-ready deployment mindset

---

🤝 Contributing

Contributions are welcome!
Please open an issue or submit a pull request for improvements or bug fixes.

---

📄 License

This project is licensed under the MIT License.

---

Built with a focus on correctness, clarity, and real-world deployment.