A new lightweight network for efficient UAV object detection is an open access work by Wei Hua, Qili Chen and Wenbai Chen.

Optimizing the structure of deep neural networks is crucial for many applications, particularly in object detection tasks for Unmanned Aerial Vehicles (UAVs). Due to the constraints of onboard platforms, more efficient networks are necessary to meet practical demands. However, existing lightweight detection networks often have excessive redundant computations and may suffer from a loss of accuracy.

This paper proposes a new lightweight network structure called Cross-Stage Partially Deformable Network (CSPDNet) to address these issues. The initial proposal features a Deformable Separable Convolution Block (DSCBlock), which significantly separates feature channels to reduce the computational load of convolution and applies adaptive sampling to the separated feature map.

A channel weighting module is also introduced to facilitate information interaction between feature layers. This module calculates weights for the separated feature map, enabling information exchange across channels and resolutions and compensating for the effects of point-wise (1×1) convolutions by filtering out the most important feature information.

Moreover, a new CSPDBlock, primarily composed of DSCBlock, has been designed to establish multidimensional feature correlations for each separated feature layer. This approach enhances the ability to capture critical feature information and reconstruct gradient paths, thereby preserving detection accuracy.

The proposed CSPDNet technology achieves a balance between model parameter size and detection accuracy. Experimental results on object detection datasets show that our designed network, with fewer parameters, achieves competitive detection performance compared to existing lightweight networks such as YOLOv5n, YOLOv6n, YOLOv8n, NanoDet, and PP-PicoDet.

The optimization effect of the designed CSPDBlock is validated using the VisDrone dataset when incorporated into advanced detection algorithms like YOLOv5m, PPYOLOEm, YOLOv7, RTMDetm, and YOLOv8m. Incorporating the designed modules resulted in a 10–20% reduction in parameters while nearly maintaining detection accuracy.

Publication Date– June 2024

A new lightweight network for efficient UAV object detection contains the following major sections:

  • Introduction
  • Related work
  • Methodology
  • Experiments
  • Conclusion

Post Image- Characteristic heat maps of the baseline model (Post Image Credit: Authors)