Automated fish identification plays a pivotal role in the development of intelligent aquaculture systems by enabling more effective stock assessment and behavioral monitoring. Although contemporary convolutional neural network (CNN)-based approaches have demonstrated strong recognition performance, they frequently exhibit computational inefficiency and limited robustness under the challenging conditions characteristic of underwater environments. In this study, we introduce a novel network exploration framework, grounded in the RegNet design paradigm, for deriving task-specific architectures tailored to underwater fish recognition. Using a relatively small dataset and approximately 200K training iterations, we obtain a family of high-performing models, collectively referred to as SeekNet, spanning multiple complexity regimes. Relative to state-of-the-art baselines, SeekNet consistently achieves superior performance. On our primary dataset, SeekNet attains a rank-1 accuracy of 95.97% and a True Acceptance Rate (TAR) of 88.04% at a False Acceptance Rate (FAR) of 1 0 − 6 . On a separate closed-set dataset, it reaches a rank-1 accuracy of 98.78% and a TAR of 98.71% at the same FAR threshold. These results substantiate the effectiveness of the proposed methodology and underscore its practical suitability for deployment in real-world aquaculture environments. • Optimized network design paradigm for superior efficiency for underwater fish recognition. • Enhanced CSP with attention and group convolution for improved performance. • Identifies high-performing models with fewer than 200k training iterations. • Achieves 95.97% Rank-1 accuracy and 88.04% TAR@10 −6 FAR on primary dataset. • Outperforms state-of-the-art models with higher accuracy and lower GPU usage.
Luo et al. (Sun,) studied this question.