Radio Signals Classification

Overview

A deep learning project for classifying radio signals by converting them to spectrogram images and applying computer vision techniques. This approach bridges signal processing and image classification.

Approach

Signal to Image Pipeline

1. Raw Signal Input: Radio frequency signals
2. Spectrogram Generation: Time-frequency representation
3. Image Classification: CNN-based classification

Data Augmentation

Implemented SpecAugment techniques for robust training:

• Time masking
• Frequency masking
• Time warping

Model Architecture

Using timm (PyTorch Image Models) for transfer learning:

• Pre-trained backbones (EfficientNet, ResNet, etc.)
• Fine-tuning on spectrogram images
• Custom classification head

Key Features

• Spectrogram Visualization: Visual representation of signal characteristics
• Transfer Learning: Leveraging pre-trained vision models
• Audio Augmentation: SpecAugment for data augmentation
• PyTorch Implementation: Clean, modular code structure

Resources

• SpecAugment Paper
• PyTorch Audio Augmentation
• timm Documentation

Links

• GitHub Repository

All projects → · The repository →