🗑️ Garbage Classification using ResNet50 & Gradio

This project implements an AI-powered garbage classification system using transfer learning with ResNet50.
The system can classify images of waste into 6 categories:

cardboard | glass | metal | paper | plastic | trash

Additionally, a Gradio-based UI is provided for easy image upload and real-time classification with top‑3 confidence scores.

---

⚡ Features

1. Deep Learning with Transfer Learning
   • Uses ResNet50 pre-trained on ImageNet.
   • Two-phase training: Feature Extraction → Fine-Tuning.
2. High Accuracy on 6 Waste Classes
   • Achieved 86.72% test accuracy with macro F1-score of 0.87.
3. Real-Time Inference
   • Gradio UI provides Top‑3 confidence predictions for new images.
4. Visual Insights
   • Confusion matrix and misclassified images for model analysis.

---

🖼️ Model Workflow

---

📊 Results

Training & Validation Performance

• Training stabilized at ~0.78–0.80 accuracy after 20 epochs
• Smooth loss curve, low overfitting due to data augmentation & early stopping

Metric	Score
Final Test Accuracy	86.72%
Macro F1-score	0.87
Weighted F1-score	0.87

---

Sample Misclassified Images

• Cardboard → Paper: Often occurs for folded or printed cardboard
• Glass → Metal: Glass jars with metallic lids reflect light like metal
• Trash: Most challenging due to few samples (22 images)

---

Inference on New Images (Top‑3 Confidence)

• High-confidence Top‑1 predictions (>89%) for clear cases
• Correct class appears in Top‑3 even in ambiguous cases
• Suitable for smart bin or waste sorting system applications

---

🏗️ How to Run

1️⃣ Clone the Repository

git clone https://github.com/YourUsername/Garbage-Classification.git
cd Garbage-Classification

2️⃣Create and activate a virtual environment (Optional):

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

3️⃣Install dependencies:

pip install -r requirements.txt

4️⃣ Run Gradio App:

python app.py

---

Gradio will launch a local server and provide a link for the web interface.

---

🧠 Lessons Learned & Challenges

1. Overfitting during Early Experiments

   • Initial models overfitted quickly due to the small dataset
   • Solution: Added data augmentation, early stopping, and fine-tuning only top ResNet blocks

2. Visual Similarity between Classes

   • Cardboard vs Paper and Glass vs Metal are the most confusing pairs

3. Small Dataset for Trash Class

   • Lower representation (22 samples) caused some misclassifications

---

✨ Future Improvements

• Add more data for underrepresented classes like Trash
• Use Progressive Fine-Tuning and Test-Time Augmentation (TTA)
• Deploy to smart bins / IoT with PyTorch Mobile or TensorRT

---

📜 License

This project is licensed under the MIT License.
Feel free to use and adapt it for academic or commercial purposes.

---

---

🌐 Author & GitHub

Developed by: Ei Khaing

📂 GitHub Repository: Click Here

---