HPRT-YOLO: Enhanced lightweight UAV maritime search and rescue model based on YOLOv11n

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Small target detection in Unmanned Aerial Vehicle (UAV) maritime search and rescue (SAR) faces challenges like low resolution, dynamic environments, and background interference.
• These issues lead to missed/false detections and inadequate real-time performance, hindering effective SAR operations.

Purpose of the Study:

• To develop an efficient and accurate lightweight model for small target detection in UAV maritime SAR.
• To address the limitations of existing methods by enhancing feature representation, multi-scale fusion, and model compression.

Main Methods:

• Proposed HPRT-YOLO, a lightweight model based on YOLOv11n, incorporating the C3k2_Faster_CGLU module for enhanced feature representation and reduced complexity.
• Strengthened multi-scale fusion using the ASF-P2 network with Zoom_cat and ScalSeq.
• Implemented LAMP pruning and BCKD knowledge distillation for parameter reduction and feature retention.
• Utilized a WIoU-v2+MPDIoU hybrid loss for optimized recall and localization.

Main Results:

• HPRT-YOLO achieved a 7% increase in mAP50 and a 9.5% increase in recall compared to YOLOv11n on public maritime SAR datasets.
• The model size was compressed by 58.19% to 2.3 MB.
• Maintained a real-time performance of 94.9 FPS at a batch size of 1.

Conclusions:

• HPRT-YOLO offers a significant improvement in small target detection for UAV maritime SAR, balancing accuracy, efficiency, and model size.
• The proposed model provides efficient technical support for real-world UAV maritime SAR applications.