Lightweight Object Detection Algorithm for UAV Aerial Imagery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• UAV aerial imagery presents challenges like low detection precision and large parameter volume due to high resolution, scale variations, and complex backgrounds.
• Existing algorithms struggle to efficiently handle these complexities, leading to suboptimal performance in object detection tasks.

Purpose of the Study:

• To develop a lightweight and precise object detection algorithm for UAV aerial imagery.
• To address challenges of scale variation and complex backgrounds in aerial image analysis.
• To improve detection accuracy and reduce computational load for real-time applications.

Main Methods:

• Introduced MFP-YOLO, a lightweight algorithm based on YOLOv5s.
• Designed a multipath inverse residual module with an attention mechanism to handle scale variations and background interference.
• Employed parallel deconvolutional spatial pyramid pooling for multi-scale target detection.
• Utilized Focal-EIoU loss function to enhance focus on high-quality samples and improve training.
• Implemented a lightweight decoupled head to accelerate convergence and boost precision.

Main Results:

• MFP-YOLO improved mAP50 by 12.9% on the VisDrone 2019 validation set and 8.0% on the test set compared to YOLOv5s.
• Reduced parameter volume by 79.2% and weight size by 73.7%.
• Demonstrated superior performance over mainstream algorithms in UAV aerial imagery detection.

Conclusions:

• MFP-YOLO effectively addresses the challenges of UAV aerial imagery detection, offering high precision with a significantly reduced model size.
• The proposed algorithm shows strong potential for real-world applications requiring efficient and accurate object detection from aerial platforms.
• MFP-YOLO represents a significant advancement in lightweight deep learning models for aerial image analysis.