Original scientific paper

https://doi.org/10.17559/TV-20240317001402

TransUNet Image 3D Reconstruction with Hyperparameter Optimization

Xiaofang Wang ; Geely University of China, Chengdu, Sichuan 641423, China *
Zhihao Luo ; Geely University of China, Chengdu, Sichuan 641423, China
Mingrui Gou ; Geely University of China, Chengdu, Sichuan 641423, China
Kerui Mao ; Geely University of China, Chengdu, Sichuan 641423, China
Liang Zhou ; Geely University of China, Chengdu, Sichuan 641423, China

* Corresponding author.

Abstract

Ancient architecture is characterized by its complexity and exquisite structure, but most existing images are in 2D format. This study proposes a TransUNet-based 3D reconstruction method with hyperparameter optimization for depth prediction to enhance the effectiveness, accuracy, and efficiency of reconstructing ancient buildings from 2D images. The method employs Restricted Boltzmann Machine (RBM) for depth prediction and an optimized ant colony algorithm for network parameter optimization. Experiments demonstrate that the proposed method achieves an average F1 score of 96.8% in reconstructing ancient buildings, outperforming other algorithms in terms of processing time and efficiency. The results validate the superiority of the proposed algorithm in processing images of ancient architecture, improving measurement accuracy and reducing execution time. This study contributes to the digitization and preservation of cultural heritage.

Keywords

hyperparameter optimization; neural networks; TransUNet; 3D Reconstruction

Hrčak ID:

330580

URI

https://hrcak.srce.hr/330580

Publication date:

1.5.2025.

Visits: 426 *