Abstract
The rapid adoption of underwater imaging systems in aquaculture has created increasing demand for automated fish monitoring technologies. Current vision-based detection methods can effectively identify and locate fish targets, but they often provide limited information about fish morphology, restricting their usefulness for applications such as stock assessment, growth evaluation, and behavioral monitoring.
To address this challenge, we propose SCALE-Net, a lightweight shape-aware Lightweight framework that combines fish detection with contour reconstruction for enhanced underwater observation. The proposed method is designed to extract representative structural features from underwater imagery and estimate anatomical landmarks that describe fish body geometry. These structural cues are subsequently used to reconstruct fish contours, enabling the generation of more informative target representations than conventional bounding-box outputs.
SCALE-Net incorporates an adaptive feature learning strategy to improve robustness under challenging underwater conditions, including low visibility, illumination variations, background interference, and partial occlusions. The lightweight architecture maintains computational efficiency while supporting accurate detection and shape estimation, making it suitable for deployment in practical aquaculture monitoring systems.
Experiments conducted on public underwater fish datasets demonstrate that the proposed framework achieves reliable fish detection performance while providing detailed contour information across diverse observation scenarios. By integrating target localization with morphological representation, SCALE-Net offers a practical tool for intelligent aquaculture management, fish population assessment, and long-term automated monitoring of aquatic environments.