Melan-YOLO: a dual-strategy framework with custom architectures for accurate and efficient underwater object detection

Effective autonomous monitoring of marine litter is vital to reduce marine pollution and protect marine ecosystems. However, current underwater detection systems face significant challenges, primarily due to poor image quality caused by light scattering in turbid waters and the limited computational resources of deployable hardware. To address these challenges, we introduce Melan-YOLO, a dual-strategy object detection framework specifically designed for underwater environments. This framework incorporates Melan-YOLO-Pro to achieve maximum detection accuracy, while offering Melan-YOLO-Edge as a streamlined alternative for deployment in resource-constrained conditions. Both architectures employ a tailored image enhancement pipeline optimized for underwater imaging and incorporate an advanced SIoU loss function to improve localization of irregularly shaped marine debris. By leveraging advanced modules such as Asymmetric Convolutions and Coordinated Attention, Melan-YOLO-Pro achieves highly competitive accuracy on the challenging RUOD and SeaClear datasets, recording 84.9% and 82.7% mAP50, respectively. Furthermore, it establishes leading localization precision with mAP50-95 scores of 61.2% on RUOD and 60.2% on SeaClear, outperforming all evaluated state-of-the-art models in this stricter metric. In parallel, Melan-YOLO-Edge significantly reduces computational demands by combining lightweight GhostConv modules with CBAM and ShuffleAttention mechanisms. It requires only 7.7 GFLOPs and 2.59M parameters, while achieving a competitive 83.6% mAP50 on RUOD alongside a 59.1% mAP50-95, and 79.7% mAP50 on SeaClear with 57.2% mAP50-95. The results demonstrate that Melan-YOLO offers a favorable balance in the accuracy-efficiency trade-off inherent in underwater object detection and provides both a high-performance and computationally efficient solution suitable for real-time applications on Autonomous Underwater Vehicles (AUVs). This framework represents a step toward practical, large-scale marine conservation and pollution monitoring, combining accuracy with computational efficiency to address real-world deployment constraints effectively.