A new lightweight network for efficient UAV object detection

Wei Hua ¹, Qili Chen ², Wenbai Chen ¹

⁰Beijing Information Science and Technology University, Beijing, 100192, China.

Scientific Reports +

|

June 10, 2024

View abstract on PubMed

Summary

A new lightweight network, Cross-Stage Partially Deformable Network (CSPDNet), enhances object detection for Unmanned Aerial Vehicles by reducing computations and maintaining accuracy. This efficient model balances parameter size and performance for practical demands.

Area Of Science

Computer Vision
Deep Learning
Artificial Intelligence

Background

Optimizing deep neural network structures is crucial for Unmanned Aerial Vehicle (UAV) object detection due to onboard platform constraints.
Existing lightweight networks often suffer from redundant computations and accuracy loss.

Purpose Of The Study

To propose a novel lightweight network, Cross-Stage Partially Deformable Network (CSPDNet), for efficient UAV object detection.
To address the trade-off between model efficiency and detection accuracy in resource-constrained environments.

Main Methods

Introduced Deformable Separable Convolution Block (DSCBlock) to reduce computational load and apply adaptive sampling.
Developed a channel weighting module for inter-feature layer information exchange and filtering.
Designed a new CSPDBlock integrating DSCBlock for multidimensional feature correlations and gradient path reconstruction.

Main Results

CSPDNet achieves a balance between model parameter size and detection accuracy.
Experimental results show competitive performance against existing lightweight networks (YOLOv5n, YOLOv8n, etc.) with fewer parameters.
Incorporating CSPDBlock into advanced models reduced parameters by 10-20% with minimal accuracy loss on the VisDrone dataset.

Conclusions

CSPDNet offers an efficient and accurate solution for object detection in UAV applications.
The proposed modules effectively reduce computational complexity while preserving critical feature information.
The CSPDNet architecture demonstrates significant potential for real-world deployment on resource-limited platforms.

Keywords:

Convolutional neural network Lightweight network Object detection Unmanned aerial vehicle (UAV)