PPDS-Net: A Physical-Prior-Guided Dual-Stream Network for Non-contact Fish Weight Estimation from Single-View 3D Point Clouds

Abstract Accurate, noninvasive estimation of fish weight is critical for biomass assessment, feeding management, and health monitoring in precision aquaculture. In underwater scenes, however, freely swimming fish undergo nonrigid deformation, adopt variable postures, and often yield sparse or noisy point clouds, which limits the robustness of existing vision-based methods. This study develops PPDS-Net, a physical-prior-guided dual-stream feature fusion network for end-to-end sturgeon weight estimation from single-view point clouds. A mask-guided single-view 3D reconstruction strategy is first used to suppress background interference and generate high-quality fish point clouds from RGB-D data. An implicit geometric stream based on point-cloud learning is then fused with an explicit physical stream composed of four morphological priors: raw point count, sampling-density ratio, spatial dispersion, and topological skeleton length. To embed morphological constraints into feature learning, the framework further introduces physical-attention modulation and physical-consistency loss. Experiments using 4,175 RGB-D samples from 86 sturgeon show that the proposed method achieves an MAE of 118.31 g and a Pearson correlation coefficient of 0.9563. Compared with the baseline PointNet + + model, the proposed framework substantially improves prediction accuracy and reduces the number of epochs required to reach the 200 g error threshold from 191 to 8. PPDS-Net also retains stable performance under highly sparse point-cloud conditions. These results demonstrate that integrating physical priors with deep point-cloud learning provides an accurate, robust, and interpretable approach for underwater fish weight estimation.