YOLO-ESFM: A multi-scale YOLO algorithm for sea surface object detection

Abstract

Environmental perception and object detection are pivotal research topics in the marine domain. The sea surface presents unique challenges, including harsh weather conditions, wave interference, and multi-scale targets, often resulting in suboptimal detection results. To address these issues, we present an innovative solution: the integration of the Efficient Scale Fusion Module (ESFM) into the advanced YOLO architecture, resulting in the enhanced model, YOLO-ESFM. The ESFM serves as both the backbone and detection head of the network, significantly improving performance compared to the baseline models in YOLOv5s, YOLOv7-tiny, and YOLOv7. Furthermore, to tackle the limitations of the CIOU in YOLOv7, we introduce an improved method, ZIOU, which has been rigorously evaluated and proven effective on the Sea Surface Target Dataset. Comparative studies demonstrate that YOLO-ESFM not only maintains efficiency in terms of parameters and FLOPs but also surpasses YOLOv7 in detection accuracy on both the Sea Surface Target Dataset and the PASCAL VOC 07+12 Dataset.