MarineYOLO: Innovative deep learning method for small target detection in underwater environments

Abstract

In the realm of underwater object detection, conventional methodologies often encounter challenges in accurately identifying and detecting small targets. These difficulties stem primarily from the intricate nature of underwater environments, suboptimal lighting conditions, and the diminutive scale of the targets themselves. To address this persistent challenge, the MarineYOLO network is introduced. This approach involves refining the conventional C2f module into the EC2f module, alongside the integration of the Efficient Multi-scale Attention (EMA) module into YOLOv8. Additionally, the Convolutional Block Attention Module (CBAM) is introduced to further refine the Feature Pyramid Network (FPN), facilitating enhanced feature extraction pertinent to small targets. Furthermore, the conventional CIoU is replaced with Wise-IoU to augment the precision and stability of target localization. Experimental findings demonstrate that MarineYOLO achieves an average precision (AP) of 78.5% on the RUOD dataset and 88.1% on the URPC dataset, marking improvements of 12.2% and 16.8%, respectively, compared to YOLOv8n. As an emerging paradigm in underwater object detection, MarineYOLO harbors significant potential in both practical applications and scholarly endeavors, furnishing an efficacious remedy to the challenges associated with detecting small targets in underwater settings.