<p>Accurate and continuous monitoring of fish body length constitutes a fundamental prerequisite for growth assessment, feed optimization, and operational decision-making in industrial recirculating aquaculture systems (RAS). However, monocular vision-based estimation in real-world underwater environments is challenged by inherent scale ambiguity, non-rigid fish locomotion, frequent pose variations, and complex optical distortions. To address these challenges, a robust estimation framework is proposed that integrates deep learning-based detection, imaging physics priors, biological morphological constraints, and temporal consistency modeling. A lightweight YOLOv11n model is employed as the frontend detector to jointly localize fish bodies and eyes within video streams, facilitating real-time deployment on constrained resource edge devices. A baseline mapping between pixel dimensions and physical length is established through reference scale calibration. Subsequently, a depth approximation method based on brightness attenuation characteristics is incorporated as a weak physical constraint to compensate for scale variations induced by anteroposterior swimming motion. Furthermore, a fish eye scale structural constraint is developed by exploiting the allometric relationship between eye diameter and body length, thereby providing reliable estimation under conditions of body deformation or incomplete segmentation. An adaptive fusion mechanism integrates these constraints according to pose confidence scores, while temporal consistency is enforced through a sliding window approach to mitigate frame-level noise and measurement fluctuations. Global linear calibration is subsequently applied at the population level to eliminate systematic scale bias. Experimental validation using datasets collected from operational aquaculture facilities demonstrates that the proposed multi-constraint framework significantly enhances the stability and robustness of monocular estimation. Statistical analysis reveals close alignment between the calibrated length distribution and manual measurements, with the Kolmogorov–Smirnov test confirming no statistically significant difference under the current experimental setting. Although single-frame measurements remain susceptible to biological motion and optical uncertainties at the individual level, the synergistic effect of multiple constraints enables a stable and approximate non-contact characterization of population growth trends. Collectively, this approach provides a pragmatic, cost-effective, and interpretable solution for fish length monitoring, offering substantial potential for advancing intelligent RAS applications.</p>