Research On Style Retrieval Technology Of 3D Model

In recent years,with the rapid growth of the 3D model design resource library,data-driven automatic design of indoor scenes has become a hot research issue in computer graphics.The thesis studies the style extraction of 3D models in the automatic design of indoor scenes.The research purpose is to retrieve 3D models with similar styles from the design resource library to drive the automatic generation of high-quality indoor scenes.The main work of the thesis is as follows:1.This paper summarizes and analyzes the research on style feature extraction of 3D models.The problem of 3D model style feature extraction involves multiple research fields such as local shape feature description,style feature extraction,and automatic generation of indoor scenes.The paper analyzes and summarizes related research,and illustrates the limitations and application prospects of existing style feature extraction studies.2.A sparse representation mechanism of 3D model style features with local shape descriptors is proposed.It is aiming at the problem that the existing algorithms cannot measure the importance of 3D local features to style features,the paper proposes a sparse representation mechanism of 3D model style features.Firstly,the algorithm obtains local features of the 3D model through a shared convolutional neural network.Secondly,the algorithm uses a Laplacian matrix to obtain a sparse representation of the local features.Finally,the importance of local features is measured by the word space histogram to obtain the global style features of the 3D model.This style feature representation not only maintains the discriminability of features through sparse learning but also solves the problem of the dimensional disaster of features.In the experimental part,the paper conducts an experimental analysis on two benchmark data,and the algorithm can achieve better retrieval performance.3.A progressive interior scene generation method is proposed to help users quickly complete interior scene furniture design.User's input is an incomplete scene and a known layout.The algorithm first retrieves style-compatible 3D models from the database.The algorithm then inserts it into the current scene,and repeats the above process until the scene design is completed.Based on the above process,the paper first improves the discriminability of style feature by defining sparse representation based convolutional network.The network uses a sparse representation to form the input matrix,optimizes the sparse features through a convolutional layer and a pooling layer.It also uses a triplet function with marginal threshold constraints to calculate the loss error.Then the thesis defines the correlation knowledge constraints for the automatic placement of three-dimensional elements and improves the interaction design efficiency of scene generation.In the experimental part,the paper compares the retrieval accuracy of the improved network structure.It can be found that the triangle extraction of style features is better than the quasi-Newton's solution strategy in Chapter 3.Finally,the thesis generates part of the indoor scene through style feature matching and three-dimensional element orientation automatic placement algorithm.