To address the issues of false detections and missed detections in underwater fish recognition under complex environmental conditions, this paper proposes FD-YOLO, a recognition method based on YOLOv8 incorporating dynamic adaptive feature learning. First, in the backbone, we design a MRFA module that combines parallel multi-scale convolution with RFA to enhance the network′s ability to capture fine-grained differences in fish features. Second, in the neck, we introduce the ECARU module, which integrates a dual-channel fusion structure with the CARAFE mechanism. This upsampling module adaptively reweights features to improve the reconstruction of local fish details. Finally, to mitigate recognition bias caused by class imbalance, we adopt Varifocal Loss, which includes a dynamic adjustment factor, to improve the accuracy of underwater fish localization. Experimental results demonstrate that, compared with YOLOv8n, the proposed FD-YOLO achieves improvements of 3.6%, 4.9% and 3.7% in Precision, Recall and mAP50.95, respectively. Moreover, the parameter size and computational cost are reduced to 2.5 MB and 6.8 GFLOPs. These findings demonstrate that the proposed method can provide technical support and reference for automated detection and monitoring of underwater targets.