FANet - Frequency-Aware Detection

FANet: Frequency-Aware Attention-Based Tiny-Object Detection in Remote Sensing Images
W Wang, Y Hu, Y Liu, Z Wang, G Zhou
Remote Sensing, 2025
DOI

Project Overview

Challenge: Tiny Object Detection in Remote Sensing

Remote sensing images contain numerous tiny objects that are difficult to detect:

• Very small object sizes in high-resolution imagery
• Similar appearance to background clutter
• Large number of objects in dense scenes
• Computational efficiency requirements

Solution: Frequency-Aware Attention Network

FANet leverages frequency domain information:

1. Frequency-aware feature extraction
2. Attention-based feature enhancement
3. Multi-scale detection head

Key Innovations

1. Frequency-Aware Feature Extraction

Utilizes frequency domain information:

• Discrete Cosine Transform (DCT) analysis
• Frequency band selection
• Spatial-frequency fusion

2. Attention-Based Enhancement

Focuses on informative regions:

• Channel attention for feature selection
• Spatial attention for location focus
• Frequency attention for band selection

3. Multi-Scale Detection

Handles objects at various scales:

• Feature pyramid network
• Scale-specific detectors
• Adaptive anchor assignment

Method

Network Architecture

Remote Sensing Image
    ↓
[Frequency Decomposition]
├── DCT transform
├── Frequency band splitting
└── Feature extraction
    ↓
[Attention Module]
├── Channel attention
├── Spatial attention
└── Frequency attention
    ↓
[Multi-Scale Detection]
├── Feature pyramid
├── Detection heads
└── Result fusion
    ↓
Detection Results

Results

Performance on remote sensing benchmarks:

• Superior tiny object detection accuracy
• Effective handling of dense scenes
• Improved computational efficiency

Applications

• Ship detection in harbors
• Vehicle detection in urban areas
• Aircraft detection at airports
• Infrastructure monitoring

Citation

@article{wang2025fanet,
  title={FANet: Frequency-Aware Attention-Based Tiny-Object Detection in Remote Sensing Images},
  author={Wang, W and Hu, Y and Liu, Y and Wang, Z and Zhou, G},
  journal={Remote Sensing},
  volume={17},
  number={2},
  pages={294},
  year={2025},
  publisher={MDPI}
}