Građevinar, Vol. 78 No. 08., 2026.
Prethodno priopćenje
https://doi.org/10.14256/JCE.4500.2026
Three-dimensional reconstruction of damaged concrete: Neural radiance field versus explicit methods
Shengpeng Zuo
Xixian Chen
Bowen Chen
Jiaqi Li
Sažetak
To explore the applicability of implicit modelling methods in the three-dimensional (3D) reconstruction of building materials, the neural radiance field (NeRF) algorithm is introduced in this paper to study the effectiveness of 3D reconstruction of images of damaged concrete surfaces, and a comparative analysis is conducted with two traditional explicit reconstruction methods, namely structure from motion and multi-view stereo. Forty-two images of concrete blocks were collected using a smartphone, and four datasets were constructed. Three-dimensional reconstruction and visual rendering were performed on these images. The results show that the peak signal-to-noise ratio of the NeRF model based on instant neural graphics primitives reaches 24.13 dB, outperforming the 17.94 dB of traditional methods. The model also exhibits higher performance in terms of modelling speed and reconstruction details. In addition, this method has obvious advantages in mitigating the surface holes and edge aliasing. The research results verify the feasibility and advantages of NeRF in the visualisation of building materials and provide a new technical path for the 3D modelling and intelligent identification of surface damage in building materials.
Ključne riječi
neural radiance field (NeRF); visualization; 3D reconstruction; intelligent recognition; deep learning
Hrčak ID:
351263
URI
Datum izdavanja:
25.9.2026.
Posjeta: 0 *