Underwater fish body detection and morphometric measurement via reliability-aware stereo vision

In digital aquaculture, fish morphometric traits, including total length (AI), body length (AH), body depth (EJ), head length (AD), eye diameter (BC), caudal peduncle length (GF), and caudal peduncle depth (LM), are important for growth monitoring, harvesting, grading, and selective breeding. However, conventional vision-based methods remain affected by underwater geometric distortion, degraded landmark visibility, and fish body deformation. To address these challenges, we propose a reliability-aware stereo vision framework. The framework employs deployment-oriented in-situ calibration within the target measurement volume to establish effective stereo geometry. Furthermore, a measurement-oriented anatomical keypoint detector integrating multi-scale dilated convolution (MSDC) and dual-path spatial-channel attention (DPSA) is developed to improve landmark localization. Subsequently, 13 anatomical landmarks are reconstructed in three dimensions to estimate the seven morphometric traits. Specifically, a quality score (Q-Score) propagates confidence-calibrated landmark uncertainty through stereo reconstruction to provide a trait-specific reliability indicator for each individual measurement. Experiments on Leopard Coral Trout and Golden Pompano show that the detector achieves 92.4% PCK@0.05, while the complete framework obtains a mean absolute error of 0.50 cm and a mean absolute percentage error of 2.78%. Q-Score is negatively correlated with actual relative error (Pearson r=−0.74 and Spearman ρ=−0.77, both p < 0.001). Retaining measurements with Q ≥ 0.90 reduces the mean absolute error of the retained measurements to 0.31 cm and limits their failure rate to 2.0% under the 5% relative-error criterion. The proposed framework provides a practical tool for reliable, non-contact multi-trait fish morphometry in controlled aquaculture operations.