SceneSense AI is a Deep Learning-based Multi-Class Image Classification project that classifies natural scene images into six categories using Transfer Learning with ResNet18.
The project was built using PyTorch for model development and Streamlit for deployment, providing an interactive web application for real-time scene classification.
🔗 Web Application:
https://scenesense-ai.streamlit.app/
🔗 GitHub Repository:
https://github.com/Vansh2639/SceneSense-AI
The project uses the Intel Image Classification Dataset, which contains thousands of images belonging to six natural scene categories:
- Buildings
- Forest
- Glacier
- Mountain
- Sea
- Street
The dataset was cleaned and validated before training to remove corrupted images.
Intel Image Classification Dataset: https://www.kaggle.com/datasets/puneet6060/intel-image-classification
- Backbone: ResNet18
- Framework: PyTorch
- Transfer Learning: Yes
- Input Size: 224 × 224
- Output Classes: 6
The final fully connected layer of ResNet18 was modified to classify the six scene categories.
The following preprocessing techniques were applied:
- Image Resizing (224 × 224)
- Random Horizontal Flip
- Random Rotation
- Tensor Conversion
- ImageNet Normalization
These transformations help improve model generalization and performance.
- Loss Function: CrossEntropyLoss
- Optimizer: Adam
- Transfer Learning with ResNet18
- GPU Training on Kaggle
The model was trained and evaluated using separate training and testing datasets.
| Metric | Value |
|---|---|
| Test Accuracy | 93% |
| Classes | 6 |
| Model | ResNet18 |
| Framework | PyTorch |
93%
- Confusion Matrix
- Classification Report
- Class-wise Precision
- Class-wise Recall
- F1 Score
The model performs well across all six scene categories and demonstrates strong generalization capability.
The trained model was deployed using Streamlit, allowing users to:
- Upload scene images
- Get real-time predictions
- View confidence scores
- View Top-3 predictions
- Explore project information through an interactive dashboard
SceneSense-AI/
├── app.py
├── intel_resnet18.pth
├── requirements.txt
├── README.md
├── notebook/
│ └── Intel_Scene_Classification.ipynb
└── Screen_shots/
├── home.png
└── prediction.png
Clone the repository:
git clone https://github.com/Vansh2639/SceneSense-AI.gitMove into the project directory:
cd SceneSense-AIInstall dependencies:
pip install -r requirements.txtRun the application:
streamlit run app.py- Python
- PyTorch
- TorchVision
- Streamlit
- NumPy
- PIL
- Scikit-Learn
- Fine-tune deeper ResNet layers
- Deploy on Streamlit Community Cloud
- Add Grad-CAM visualizations
- Experiment with EfficientNet and Vision Transformers
- Add batch image prediction support
Vansh Garg
B.Tech CSE
Machine Learning & AI Enthusiast
⭐ If you found this project useful, consider giving it a star.