In order to achieve high-precision and automatic measurement of the phenotype of carpenters, and to solve the problems of insufficient phenotypic data of two-dimensional image processing and insufficient recognition ability of existing three-dimensional point cloud segmentation models for complex biological structures, this study proposed a three-dimensional morphological phenotypic extraction method based on end-to-end point cloud segmentation. In this method, the multi-view image sequence collected by RGB camera is used to reconstruct a high-fidelity 3D point cloud model by motion recovery structure (SfM) and multi-view stereo vision (MVS) algorithms. In order to effectively separate the target from the background, a composite denoising method combining the random sampling consistency (RANSAC) algorithm and HSV spatial color features is designed to accurately remove background noise while preserving the details of the fish body. The experimental results showed that the MSE was lower than 10e-5, the average PSNR was 60.8, and the average recall was 93.33%. Aiming at the problem of insufficient feature extraction of the original PointNet++ network, the segmentation performance of the fine structure of the fish body is significantly improved by introducing a multi-layer perceptron (MLP) based on dense connection, a triple attention mechanism and an enhanced extraction module for color information. The improved network model has 88.9%, 95.5% and 93.8% of the average intersection and union ratio (MIoU), overall accuracy (OA) and average accuracy (mAcc), respectively, which are 3.5%, 1.1% and 1.6% higher than the original model. It can accurately segment key components such as fish head, fish body and fins. This study provides an effective technical solution for high-throughput, non-contact accurate measurement of aquatic animal phenotypes.