Abstract Underwater object detection plays a vital role in marine science, ecological monitoring, and underwater robotics. However, the challenging nature of underwater imagery, characterized by low visibility, poor lighting, turbidity, and texture degradation, makes reliable object detection difficult, particularly on resource constrained hardware. With the growing need for deployable and efficient solutions, this work centers around optimizing lightweight YOLO models for stable and accurate performance under underwater settings. This work analyzes the performance of four lightweight YOLO models, namely YOLOv8n, YOLOv9t, YOLOv11n, and YOLOv12n upon a labeled dataset of underwater data with object classes like echinus, holothurian, scallop, and starfish. The models were trained by precisely managed experiments where the parameter settings like learning rate, momentum, learning rate scheduling, and other regularization methods were varied step by step. The objective was to improve detection accuracy and model robustness without introducing any additional computational overhead. This systematic tuning process provided justice-based, consistent and reproducible evaluations while retaining the models lightweight for real-time execution on hardware-constrained devices. All the models provided good detection performance, each excelled under varying settings. Out of the lot, YOLOv11n had the most impressive detection accuracy, achieving an mAP@50 of 0.848 and an mAP@50:95 of 0.524 on the test set. It also achieved the maximum Recall of 0.783, emphasizing its capability to retain stable object detection over all the varied settings over the set of IoU thresholds. Class-wise observation also validated that YOLOv11n competitive and stable performance among the evaluated lightweight models. Although the YOLOv11n was the most stable across most aspects, YOLOv9t had better precision under some categories, whereas the models YOLOv8n and YOLOv12n had better compromise between speed over accuracy. Those observations verify that given the precise optimization, even the lightweight neural networks have the potential to outperform even better under the task of underwater object detection, thereby open up the possibility for low-power, real-world marine-based applications.
Ramyashree et al. (Sun,) studied this question.