HFC-UAVDet: Lightweight small object detection network for UAV with high-frequency perception and cross-layer...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Visible light images suffer from complex backgrounds, while infrared images have low contrast and weak textures.
• Small objects with sparse features in aerial imagery pose challenges for lightweight detection models.
• Balancing accuracy and efficiency in object detection for unmanned aerial vehicles (UAVs) is crucial.

Purpose of the Study:

• To propose HFC-UAVDet, a lightweight small object detection network for UAVs.
• To improve the detection of small objects in both visible light and infrared aerial imagery.
• To enhance the balance between accuracy and efficiency in UAV-based object detection.

Main Methods:

• Introduced the Cross-High-Frequency Fusion Block (CHFB) for noise suppression and edge enhancement.
• Developed the High-Resolution Dynamic Position encoder (HRDP) for stable localization in dense scenes.
• Implemented the Semantic-Guided Cross-layer Fusion (SGCF) module for aligning multi-level features.
• Integrated a P2 branch for improved ultra-small target representation.

Main Results:

• HFC-UAVDet demonstrated significant improvements in mean Average Precision (mAP50) on VisDrone2019, HIT-UAV, and TinyPerson datasets.
• Achieved performance gains of 6.6%, 2.7%, and 2.2% over RT-DETR, respectively.
• Maintained a lightweight model size of 10.7MB parameters.

Conclusions:

• HFC-UAVDet effectively addresses the challenges of small object detection in diverse aerial imagery.
• The proposed network offers a superior balance of accuracy and efficiency for UAV applications.
• The integration of high-frequency perception and cross-layer semantic fusion proves beneficial for minute object detection.