A Lightweight YOLO Model With Object Detection for Marine Aerial Imagery

Aerial image object detection in open-water regions encounter numerous difficulties. Objects typically appear tiny, leading to sub-optimal detection precision. Additionally, existing algorithms often suffer from many network parameters, resulting in high computational complexity. Current research efforts focus on developing algorithms that can achieve a balance between accuracy and speed. These algorithms also need to be suitable for deployment on small-scale mobile devices, ensuring practicality in real-world applications. A lightweight small object detection network RDU-YOLO based on YOLO algorithm is proposed. Firstly, To enhance model efficiency, we introduce the Dynamic Inception Mixer module, which addresses traditional convolution limitations. By employing shared convolution, we achieve a lightweight model. However, recognizing that shared parameters might limit feature differentiation, we implement Rep-convolution to boost feature extraction and counteract precision losses. Additionally, a bidirectional feature pyramid neck network and optimized upsampling operations further improve edge computing efficiency. These innovations maintain high detection accuracy while reducing computational overhead and parameter count. The experimental results on the SeaDroneSee dataset show that compared with the original YOLO, RDU-YOLO achieves a remarkable reduction in both parameter count and computational overhead while preserving comparable mean average precision (mAP) performance. This advancement presents a novel framework for enabling unmanned aerial systems to execute object detection missions in open water scenarios with enhanced computational efficiency and real-time capability.