Fish are an important source of dietary protein for humans, contributing significantly to national food security and public health. The body size of fish holds significant guiding and practical value for both aquaculture and marine fishing industries. With the advancement of computer vision, non-contact measurement methods are gradually replacing traditional labor-intensive manual measurements to acquire fish morphological characteristics. However, current computer vision methods cannot construct complete three-dimensional models of fish, failing to meet the current demand in the aquaculture sector for three-dimensional digital models of fish. this study utilized structured light projection combined with binocular stereovision methods to reconstruct three-dimensional digital models of fish. A fish point cloud acquisition system was designed, incorporating deep learning networks for fish body image segmentation during the acquisition process. Phase-shifting was used to mark the grayscale on the surface of fish bodies, and finally, a binocular stereovision system was employed to reconstruct fish point clouds. The accuracy of the system was validated using the economically important mackerel species (Scomberomorus niphonius). Results indicated that this method could construct fish point cloud models, with relative errors for fork length, body height, maximum body circumference, and posterior gill cover circumference of the reconstructed mackerel models being 0.82%, 4.47%, 3.14% and 2.87%, respectively. Correlation analysis and fitting of the relationships between body measurements showed that the multivariate linear regression method (R²=0.826/0.833) outperformed linear regression methods. This study provides a methodological reference for digital information collection in the fisheries industry.