Research On Parametric Modeling Method Of 3D Fish Body Based On Instance Data

Due to the dim underwater environment caused by the refraction and scattering of light,it is difficult to collect 2D and 3D data of fish,resulting in slow progress in many automated researches driven by fish data.With the wide application of 3D laser scanners,it is possible to acquire high-precision 3D fish models,but batch acquisition is still labor-intensive,time-consuming and costly.Based on the small-scale 3D fish data set collected by 3D scanners,this paper proposes two parametric modeling methods for 3D fish body shapes.By constructing a parametric representation model and changing the weight parameters,a large number of 3D fish models of different body shapes can be quickly generated,which can provide data support for researches such as fish growth prediction and fish density estimation based on 3D data.First,a set of fish data collection scheme is designed,which collects 2D images and 3D grid data of fish,and builds a standard template based on the collected data.Select grass carp with obvious changes in body size during the growth process as the target fish for parametric modeling,select a medium-sized grass carp to collect its multi-view images,and design and build a 3D fish scanning data acquisition platform to collect small-scale fish mesh models of different body sizes;based on the collected 2D images and 3D meshes,a 3D fish standard template is constructed in 3ds Max.Second,a 3D fish body mesh registration method is proposed to construct a 3D fish body same topology mesh dataset.The number of vertices and triangles of the 3D fish mesh model collected by the scanner is huge,and the topology is messy,which cannot be put into computer statistical analysis and modeling in batches.This paper designs a three-module grid non-rigid registration method based on the NRICP algorithm:the pre-processing module rigid registration provides the non-rigid registration with the initial state of the registration between the models through coordinate space transformation and scaling transformation;the non-rigid registration is based on the NRICP algorithm,and the feature marker items are not used in the objective function to achieve high-precision and high-efficiency registration;the post-processing module fin triangular face penetration detection and stretching solve the problem of self-penetration of thin fins.Finally,two parametric modeling methods of 3D fish body shape based on principal component analysis and VAE neural network are proposed to statistically analyze and model 3D fish body models of different body shapes in the same topology 3D fish body dataset.The first is to use the classical principal component analysis method to extract body shape features to build a 3D fish body parameterized representation model,and model the fins in blocks,and quickly generate a 3D1 fish body model of different body sizes by applying a set of weight parameters to the feature vector.The second is to build an unsupervised generative VAE neural network for parametric modeling of 3D fish body shapes.The neural network is successfully used in parametric modeling to quickly generate 3D fish body models of different body sizes.Among them,an interactive parametric modeling system is constructed based on principal component analysis to generate a 3D fish model in real time.