🖼️ Scene Sorter

AI-Powered Scene Classification & Smart Image Organizer

Scene Sorter is an open-source, end-to-end AI tool that automatically classifies and organizes images into real-world scene categories such as mountain, forest, glacier, sea, street, and buildings.

It is designed for travelers, photographers, dataset curators, and AI engineers who want their images clean, structured, and instantly usable.

Built with a production-ready mindset — from model training → API → UI → deployment.

---

🌍 Live Project

• Frontend (UI): https://-FRONTEND-LINK (will upate soon when deployed)

• Backend API: https:-BACKEND-LINK (will upate soon when deployed)

• Swagger Docs: https:/-BACKEND-LINK/docs (will upate soon when deployed)

---

📸 Demo & Screenshots

(Add images later — paths prepared)

Upload Images	Batch Results

ZIP Download	Swagger API

---

❓ The Real Problem (Why I Built This)

As someone who travels a lot, I constantly faced this problem:

• After trips, I had hundreds of photos
• Mountains, glaciers, streets, seas — all mixed
• Manual sorting is slow, boring, and error-prone
• Existing gallery apps don’t offer semantic scene sorting

There was no simple tool that could:

• Understand what is in an image
• Group images meaningfully
• Export them cleanly for reuse

So instead of accepting the pain, I built Scene Sorter.

This project is a problem-solver’s answer, not just a model demo.

---

🧠 Model Journey (From Scratch → Production)

Phase 1 — CNN From Scratch

• Built a custom CNN architecture
• Trained on scene classification dataset
• Accuracy achieved: ~75%
• Problems:
  • Overfitting
  • Slower convergence
  • Limited generalization

➡️ Good learning phase, but not production-grade

---

Phase 2 — Transfer Learning (Breakthrough)

Switched to MobileNetV2 (ImageNet weights):

• Used pretrained feature extractor
• Added custom classification head
• Fine-tuned higher layers
• Applied dropout & regularization

📈 Results:

• Accuracy: ~86–87%
• Faster convergence
• Better generalization
• Lower inference latency

This is the model now running in production.

---

🏷️ Scene Classes

["buildings", "forest", "glacier", "mountain", "sea", "street"]

Each prediction returns:

• label
• confidence score
• full probability vector

---

⚙️ How the System Works

High-Level Flow

User Uploads Images
        ↓
Frontend (Next.js)
        ↓
FastAPI Backend
        ↓
Image Preprocessing (PIL + NumPy)
        ↓
TensorFlow Model (MobileNetV2)
        ↓
Class-Based Folder Sorting
        ↓
ZIP File Generation
        ↓
Download Link Returned

---

🧩 Key Features

🔹 Single Image Prediction

• Upload one image

• Get:

  • predicted label
  • confidence score

🔹 Batch Image Classification

• Upload multiple images at once
• Model runs batch inference
• Images grouped by predicted class

🔹 Automatic Folder Sorting

output/
├── mountain/
│   ├── img1.jpg
│   └── img2.jpg
├── glacier/
└── forest/

🔹 ZIP Export

• Entire sorted directory is compressed
• Downloadable as a single ZIP

---

🛠️ Tech Stack

Frontend

• Next.js 14
• TypeScript
• Tailwind CSS
• Drag-and-drop uploads
• Batch preview & progress UI

Backend

• FastAPI
• Python 3.11
• Uvicorn
• RESTful architecture
• Swagger documentation

Machine Learning

• TensorFlow / Keras
• MobileNetV2
• Transfer Learning
• Softmax classification
• Batch inference optimization

DevOps & Deployment

• Docker
• Render
• GitHub CI workflow
• Environment-based config

---

🧪 Testing the System

1️⃣ Single Image Test

• Endpoint: POST /predict
• Upload one image
• Returns:

{
  "filename": "mountain.jpg",
  "label": "mountain",
  "confidence": 0.82
}

---

2️⃣ Batch Test

• Endpoint: POST /predict/batch
• Upload multiple images
• Returns:
• per-image predictions
• ZIP download link

---

🚀 Deployment Strategy

• Backend deployed as Docker Web Service on Render

• Frontend deployed separately

• Environment variables control:

  • model path
  • CORS
  • batch limits

• CPU-only deployment (no GPU required)

---

📈 Performance

• Accuracy: ~86–87%
• Latency: Optimized via lightweight MobileNetV2
• Scalability: Batch-ready, stateless API
• Portability: Dockerized & cloud-ready

---

🔮 Future Improvements (Vision)

This project is intentionally built to grow.

Planned Enhancements

• 📱 Mobile integration

  • Auto-sort photos directly on device
  • No manual download required

• 🧠 Multi-model support (ViT, CLIP)

• ☁️ Cloud storage integration

• 👤 User accounts & personal galleries

• 📊 Analytics dashboard

• 🎯 Custom class training via UI

Ultimate goal: Photos auto-organize themselves the moment you return from a trip.

---

🤝 Open Source & Contributions

This project is fully open-source.

You can:

• Improve the model
• Add new classes
• Optimize inference
• Enhance UI/UX
• Extend to mobile

PRs, issues, and ideas are welcome.

---

📜 License

MIT License — free to use, modify, and distribute.

---

👋 Author

Built by Parnish AI Engineer • ML Practitioner • Full-Stack Developer

“A good AI project doesn’t stop at training — it ships.”

⭐ If this project impressed you, consider starring it.

---