Visual feature enhancement method based on 3D convolutional network for automatic virtual design

Read full text Summarise with VIRAT
Authors: Zhengdao Gu, Bin Wang, Shaoqing Wang
Cited by: 0
References: 20

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Deep learning with 3D representation has created new opportunities for the automation of virtual design, but obtaining descriptive and ready-to-design visual features from point clouds is still difficult. To help solve this issue, this paper presents an Adaptive Multi-scale Geometric Feature Enhancement (AMGFE) approach using a 3D convolutional neural network (CNN) based architecture. This work combines coarse-scale contextual information from voxel grids with fine-scale geometrical information from raw point clouds, which are further enhanced by employing a geometry-aware adaptive convolutional mechanism that adjusts extracted features dynamically based on local shape properties. In addition, in order to more effectively highlight design-relevant characteristics such as edge and surface features, a channel-wise attention module is used. The AMGFE approach is tested against the ModelNet40 benchmark; the approach has an overall accuracy rate (OA) of 94.38%, a mean class accuracy (mCA) of 92.17%, and a macro F1 score of 0.9139. The AMGFE process outperformed state-of-the-art approaches including PointNet and DGCNN. Overall, the results of this study show that the AMGFE method balances global structural understanding with local geometrical accuracy in a manner that makes it a good candidate for downstream applications in virtual design related activities such as mesh creation and 3D modeling.

Fields of Study & Subjects

3D Shape Modeling and Analysis Advanced Neural Network Applications Image Processing and 3D Reconstruction

Publication details

Journal / source
Egyptian Informatics Journal
Volume
34
Pages
100925–100925
Publication date
25 March 2026
Language
EN