RFVision is a deep learning project focused on automatic wireless signal classification using Radio Frequency (RF) spectrogram images. The project investigates how modern computer vision architectures can learn discriminative patterns from RF spectrum representations and accurately classify different wireless signal types.
Unlike traditional RF signal processing approaches that rely heavily on handcrafted features, RFVision leverages deep learning models to automatically learn hierarchical representations directly from spectrogram images.
The project benchmarks multiple state-of-the-art architectures including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), Hybrid CNN-Transformer models, and class-imbalance-aware learning techniques to identify the most effective approach for RF signal recognition.
Wireless communication systems generate complex RF signals that vary across different technologies and transmission protocols. Identifying and classifying these signals is an important task in:
- Spectrum monitoring
- Cognitive radio systems
- Wireless network management
- Signal intelligence
- Interference detection
- Automated spectrum analysis
Traditional machine learning methods often require extensive feature engineering. This project explores whether deep learning models can directly learn meaningful patterns from spectrogram images and improve classification performance.
This project uses the Radio Frequency Signal Image Classification Dataset available on Kaggle.
https://www.kaggle.com/datasets/halcy0nic/radio-frequecy-rf-signal-image-classification
- RF spectrogram image dataset
- Multi-class classification problem
- 21 wireless signal classes
- Image-based representation of RF spectrum activity
- Suitable for CNN and Vision Transformer architectures
Each sample represents a spectrogram image generated from RF signal observations. These images contain frequency and time-domain information that can be used for signal identification.
Complete implementation is available on Kaggle:
https://www.kaggle.com/code/hetmonpara0503/wireless-signals
The primary goals of RFVision are:
- Classify wireless signals using deep learning
- Compare CNN and Transformer-based architectures
- Evaluate transfer learning approaches
- Address class imbalance issues
- Improve model interpretability using Explainable AI
- Benchmark multiple architectures on a common dataset
The project follows a complete deep learning workflow.
- Dataset inspection
- Class distribution analysis
- Sample visualization
- Dataset quality verification
- Image loading
- Image resizing
- Normalization
- Data augmentation
- Train-validation-test preparation
Multiple deep learning architectures are implemented and evaluated.
A residual convolutional neural network used as a strong baseline model.
Advantages:
- Deep feature extraction
- Residual learning
- Proven performance on image classification tasks
A computationally efficient CNN architecture that scales network dimensions effectively.
Advantages:
- Better parameter efficiency
- Strong feature representation
- Faster training
Transformer-based architecture adapted for image classification.
Advantages:
- Global attention mechanism
- Long-range dependency learning
- State-of-the-art visual understanding
Combines CNN feature extraction with transformer-based attention mechanisms.
Advantages:
- Local feature learning
- Global context modeling
- Improved representation capability
Advanced learning framework designed to handle class imbalance.
Advantages:
- Better minority class recognition
- Improved class separation
- More robust feature embeddings
To improve model interpretability, Grad-CAM is incorporated into the workflow.
Grad-CAM helps visualize:
- Important image regions
- Model attention areas
- Classification reasoning
- Feature importance
This improves transparency and trustworthiness of model predictions.
Models are evaluated using standard classification metrics:
Overall classification performance.
Measures prediction correctness.
Measures ability to identify all relevant samples.
Balances precision and recall.
Provides class-wise performance insights.
- Python
- PyTorch
- Torchvision
- TIMM
- NumPy
- Pandas
- Matplotlib
- Seaborn
- Scikit-Learn
- Grad-CAM
- Kaggle Notebook Environment
RF Spectrogram Images
│
▼
Data Preprocessing
│
▼
Data Augmentation
│
▼
Model Training
│
├── ResNet50
├── EfficientNet
├── Vision Transformer
├── CNN-ViT Hybrid
└── LDAM + SupCon
│
▼
Performance Evaluation
│
▼
Explainability Analysis
│
▼
Model Comparison
- RF signal classification from spectrogram images
- Deep learning-based feature extraction
- Transfer learning implementation
- CNN model benchmarking
- Vision Transformer benchmarking
- Hybrid architecture evaluation
- Class imbalance handling
- Grad-CAM explainability
- Comparative model analysis
- End-to-end Kaggle implementation
RFVision/
│
├── RFVision_Wireless_Signal_Classification.ipynb
├── README.md
├── LICENSE
└── .gitignore
Potential future extensions include:
- Real-time RF signal classification
- RF anomaly detection
- Self-supervised representation learning
- Multi-label signal recognition
- Deployment as a web application
- Edge-device optimization
- Real-time spectrum monitoring dashboard
RFVision can be applied in:
- Wireless communication systems
- Cognitive radio networks
- Spectrum monitoring
- Signal intelligence
- Telecommunications research
- RF security systems
- Automated spectrum analysis
Het Monpara
M.Tech ICT (Machine Learning)
Dhirubhai Ambani University
This project is released under the MIT License.# RFVision