Light-YOLO: a lightweight detection algorithm based on multi-scale feature enhancement for infrared small ship target

Abstract

Infrared-based detection of small targets on ships is crucial for ensuring navigation safety and effective maritime traffic management. However, existing ship target detection models often encounter missed detections and struggle to achieve both high accuracy and real-time performance at the same time. Addressing these challenges, this study presents Light-YOLO, a lightweight model for ship small target detection. Within the YOLOv8 network architecture, Light-YOLO replaces conventional convolutions with snake convolutions, effectively addressing the issue of inadequate detection point receptive fields for small targets, thereby enhancing their detection. Additionally, a Multi-Scale Feature Enhancement Module (MFEB) is introduced to refine focus on low-level features through multi-scale and selection strategies, mitigating issues such as interference from image backgrounds and noise during small target detection. Furthermore, a novel loss function is designed to dynamically adjust the proportions of its components during training, improving the regression accuracy of small targets towards real annotation boxes and enhancing the localization ability of detection boxes. Experimental results demonstrate that Light-YOLO outperforms YOLOv8n, achieving optimal performance on an infrared ship small target detection dataset with 9.2G FLOPs. It notably enhances accuracy, recall rate, and average precision by 1.76%, 0.83%, and 2.27%, respectively.