This study investigates the applicability of a deep learning–based multi-scale image segmentation framework for underwater biological recognition in the Chinese context. The research is grounded in controlled experimental evaluations conducted using publicly available benchmark datasets of underwater organisms. The experimental results demonstrate that the proposed multi-scale segmentation approach outperforms conventional segmentation methods by more effectively capturing underwater biological targets in complex and heterogeneous environments. By leveraging multi-scale feature representation, the model exhibits a strong capability to identify organisms of varying shapes and sizes under challenging underwater conditions. Quantitative evaluation indicates that the proposed method achieves an overall F1-score of 90.7%, highlighting its effectiveness in pixel-level segmentation tasks for underwater biological recognition. The analysis further reveals that underwater recognition remains challenging due to substantial scale variability among biological targets, with certain species occupying less than 5% of the image area, while larger organisms such as fish may dominate a significant portion of the visual field. Despite these challenges, robustness assessments confirm that the multi-scale segmentation model maintains stable and high recognition performance across different scale categories. Specifically, recognition accuracy reaches 87.6% for small-scale targets and exceeds 94% for large-scale organisms. Overall, the findings demonstrate that deep learning–based multi-scale image segmentation offers significant potential for accurate identification and monitoring of underwater biological species in diverse and dynamic environments. The proposed approach provides a robust and scalable solution for underwater ecosystem monitoring, particularly in scenarios characterised by complex illumination conditions, background interference, scale diversity, and environmental variability.
