Re-Parameterization After Pruning: Lightweight Algorithm Based on UAV Remote Sensing Target Detection

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Lightweight object detection is crucial for unmanned aerial vehicles (UAVs) in remote sensing.
• Existing algorithms often trade detection accuracy for speed, which is insufficient for UAV applications.
• There's a need for algorithms that offer both high inference speed and precise detection accuracy.

Purpose of the Study:

• To develop a refined lightweight object detection algorithm for UAV remote sensing.
• To enhance both the inference speed and detection accuracy of existing lightweight models.
• To address the limitations of current algorithms in terms of precision.

Main Methods:

• Utilized MobileNetV3 with an efficient channel attention (ECA) module as the backbone.
• Incorporated Focal and Efficient Intersection over Union (FocalEIoU) loss for improved regression and reduced false negatives.
• Applied convolution kernel pruning to reduce model parameters and FLOPs.
• Implemented a plug-and-play structural re-parameterization fine-tuning method.

Main Results:

• Model parameters and FLOPs were reduced to 1.2M/6.2G (VisDrone) and 1.5M/6.5G (DIOR) post-pruning.
• Achieved inference speeds of 49 FPS (VisDrone) and 44 FPS (DIOR) on Jetson AGX Xavier.
• Fine-tuning improved mAP@0.5 by 0.4% (VisDrone) and mAP@0.5:0.95 by 0.5% (DIOR).
• Outperformed mainstream lightweight algorithms in parameters, FLOPs, and mAP metrics, with notable reductions in missed and duplicate detections.

Conclusions:

• The proposed refined lightweight object detection algorithm offers a superior balance of speed and accuracy for UAV remote sensing.
• The combination of ECA, FocalEIoU, pruning, and re-parameterization effectively enhances detection performance.
• The algorithm demonstrates practical utility by significantly minimizing detection errors in real-world validation tests.