Article: 3D visual generation system based on the fusion of multi-view geometry reconstruction and deep learning Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.71 pp.1 - 25 Abstract: A 3D vision generation system combining multi-view geometric reconstruction and deep learning is proposed to improve robustness, reconstruction completeness, and rendering realism in complex scenes with weak texture, occlusion, and lighting variation. A cross-attention neural matcher enhances feature matching, a geometry-constrained cost volume network improves dense reconstruction, and a geometry-guided NeRF optimisation increases rendering realism. Geometry-aware attention reduces occlusion and repetitive texture interference, while sparse geometric priors improve weak-texture regions and unify geometric consistency with visual realism. Experiments show 91.7% matching accuracy and 84.2% recall, with pose errors of 0.19° and 1.14 cm. Dense reconstruction achieves 0.278 mm accuracy and 0.362 mm completeness on DTU, and 79.88 F-score on tanks and temples. Rendering reaches PSNR 34.72, SSIM 0.971, and LPIPS 0.039. The system enables robust, realistic end-to-end 3D generation from mobile image sequences for autonomous driving, industrial modelling, and digital twins. Inderscience Publishers - linking academia, business and industry through research

Title: 3D visual generation system based on the fusion of multi-view geometry reconstruction and deep learning

Authors: Guofeng Hu; Yan Lin

Addresses: College of Art and Design, Zhanjiang University of Science and Technology, Zhanjiang, 524084, China ' College of Art and Design, Zhanjiang University of Science and Technology, Zhanjiang, 524084, China

Abstract: A 3D vision generation system combining multi-view geometric reconstruction and deep learning is proposed to improve robustness, reconstruction completeness, and rendering realism in complex scenes with weak texture, occlusion, and lighting variation. A cross-attention neural matcher enhances feature matching, a geometry-constrained cost volume network improves dense reconstruction, and a geometry-guided NeRF optimisation increases rendering realism. Geometry-aware attention reduces occlusion and repetitive texture interference, while sparse geometric priors improve weak-texture regions and unify geometric consistency with visual realism. Experiments show 91.7% matching accuracy and 84.2% recall, with pose errors of 0.19° and 1.14 cm. Dense reconstruction achieves 0.278 mm accuracy and 0.362 mm completeness on DTU, and 79.88 F-score on tanks and temples. Rendering reaches PSNR 34.72, SSIM 0.971, and LPIPS 0.039. The system enables robust, realistic end-to-end 3D generation from mobile image sequences for autonomous driving, industrial modelling, and digital twins.

Keywords: feature matching; dense reconstruction; geometric reconstruction; multi-view; deep learning; DL.

DOI: 10.1504/IJICT.2026.154473

International Journal of Information and Communication Technology, 2026 Vol.27 No.71, pp.1 - 25

Received: 23 Oct 2025
Accepted: 02 Mar 2026
Published online: 29 Jun 2026 *

Title: 3D visual generation system based on the fusion of multi-view geometry reconstruction and deep learning

Keep up-to-date