Computer vision advances enable aquatic resource monitoring, but underwater environments pose challenges such as low light, blurred images, and object overlap. This study integrates YOLOv8 and RF-DETR with SAM 2.1 for underwater fish detection and segmentation. Using 8,242 images across 13 fish classes, the models were optimized through learning rate tuning, data augmentation, and memory-based video processing. YOLOv8 + SAM 2.1 delivers high performance with fast inference suitable for resource-constrained devices, while RF-DETR + SAM 2.1 achieves higher accuracy (mAP@0.5 = 82% at lr = 2e−5) in complex overlapping scenarios. In addition, a curated 700-image segmentation subset was constructed for evaluation, offering potential as a benchmark for future underwater vision research.