Skip to main content

Architecture

This document describes the architecture and general flow of the Spectrum Analyzer system.

Repository: GitHub

System Overview​

The Spectrum Analyzer uses a modified YOLO (You Only Look Once) architecture based on the EMS-YOLO framework. The system processes spectrogram images to detect and classify various signal types.

General Flow​

Input Spectrogram Image (512×512)
    ↓
Preprocessing & Normalization
    ↓
Time Window Processing (5 frames)
    ↓
ResNet Backbone (Feature Extraction)
    ↓
Multi-scale Feature Maps
    ↓
YOLO Detection Head
    ↓
Bounding Box Predictions + Class Scores
    ↓
Non-Maximum Suppression (NMS)
    ↓
Output: Detected Signals with Bounding Boxes

Architecture Components​

1. Input Processing​

Image Input:

  • Default resolution: 512×512 pixels
  • Channels: 3 (RGB spectrogram representation)
  • Configurable via --imgsz parameter

Time Window Integration:

  • Processes 5-frame temporal windows
  • Each frame is duplicated to create temporal context
  • Helps capture temporal signal characteristics

2. Backbone Network​

The backbone extracts features from input images. Supported architectures:

ResNet-18 (Default):

  • BasicBlock layers for feature extraction
  • Multi-scale feature maps at different resolutions
  • Configurable via model YAML files

Other Options:

  • ResNet-10: Lighter model
  • ResNet-34: Deeper model
  • Custom architectures via YAML configuration

Model architecture details are available in the training output directory.

3. Detection Head​

YOLO Detection Layers:

  • Two detection scales:
    • P4/16: Medium-scale detection (stride 16)
    • P5/32: Large-scale detection (stride 32)
  • Each layer outputs bounding boxes, confidence, and class predictions

Anchor Boxes:

  • Predefined anchor boxes optimized for signal detection
  • P4/16 anchors: [10,14, 23,27, 37,58]
  • P5/32 anchors: [81,82, 135,169, 344,319]
  • Configurable in model YAML files

4. Output Processing​

Detection Output:

  • Bounding boxes (x, y, width, height)
  • Confidence scores (0.0-1.0)
  • Class predictions (11 classes for RadDet)

Post-processing:

  • Non-Maximum Suppression (NMS) to remove duplicate detections
  • Confidence threshold filtering
  • Class-specific filtering (optional)

Model Configuration​

Models are configured via YAML files in the models/ directory:

# Example: models/resnet18.yaml
nc: 11  # number of classes
backbone:
  [[-1, 1, Conv_1, [64, 7, 2]],
   [-1, 1, BasicBlock_2, [64, 3, 2]],
   # ... more layers
  ]
head:
  [[-1, 1, Detect, [nc, anchors]],]

Time Window Processing​

The system uses a 5-frame time window for temporal context:

  1. Input image is duplicated 5 times
  2. Each frame is processed through the backbone
  3. Features are aggregated across time dimension
  4. Final detection uses temporal information

This helps capture:

  • Signal continuity over time
  • Temporal patterns in spectrograms
  • Better detection of time-varying signals

Feature Extraction Flow​

Input Image (1, 3, 512, 512)
    ↓
Time Window Expansion (5, 3, 512, 512)
    ↓
ResNet Backbone
    ├─→ Conv + BasicBlock layers
    ├─→ Multi-scale features
    └─→ Feature maps at different resolutions
    ↓
Detection Layers
    ├─→ P4/16 features → Medium detections
    └─→ P5/32 features → Large detections
    ↓
Temporal Aggregation (sum over time dimension)
    ↓
Final Predictions

Detection Process​

  1. Feature Extraction: Backbone extracts multi-scale features
  2. Detection Prediction: YOLO head predicts boxes, confidence, classes
  3. Temporal Aggregation: Features aggregated across time window
  4. NMS: Non-maximum suppression removes duplicates
  5. Filtering: Confidence and class filtering
  6. Output: Final detections with bounding boxes

Performance Characteristics​

Inference Speed:

  • Depends on image size and model architecture
  • ResNet-18: ~XX ms per image (512×512)
  • Configurable via batch size and device selection

Memory Usage:

  • Model size: ~XX MB (ResNet-18)
  • GPU memory: Depends on batch size
  • Configurable via quantization

Customization​

The architecture can be customized:

  1. Model Backbone: Change in model YAML file
  2. Image Size: Configure via --imgsz parameter
  3. Number of Classes: Set in dataset YAML file
  4. Anchors: Modify in model YAML file
  5. Time Window: Adjust in models/yolo.py (time_window variable)

For detailed configuration options, see Configuration Guide.