Buy article PDF
The purchased file will be sent to you
via email after the payment is completed.
US$ 35
|
Smart Structures and Systems Volume 38, Number 2, August 2026 , pages 205-228 DOI: https://doi.org/10.12989/sss.2026.38.2.05 |
|
|
|
Geometry-constrained super-resolution architecture for large field-of-view binocular camera calibration |
||
Lizhi Long, Weiqi Mao, Lu Deng, Yuanfeng Duan, Shuo Wang
|
||
| Abstract | ||
| For vision-guided automatic assembly construction robots, binocular camera calibration is essential for accurate three-dimensional perception and spatial localization. However, large field-of-view (FOV) camera systems often suffer from degraded calibration accuracy due to low-resolution and blurred checkerboard features captured at long working distances. Although conventional single-image super-resolution (SISR) methods improve image quality, they may introduce structural distortions that compromise geometric fidelity and reduce calibration accuracy. To address this issue, a Geometrically Constrained Super-Resolution Reconstruction Network (GC-SRNet) is proposed for geometry-preserving image reconstruction in camera calibration. The proposed framework integrates geometric priors with dual-level geometric supervision strategy. Specifically, the Geometric Feature Awareness Module (GFAM) and Geometric Constraint Refinement Module (GCRM) are developed to extract and refine geometry-sensitive features, while the supervision mechanism preserves geometric consistency during reconstruction. In addition, a calibration-oriented evaluation metric, termed TSR, is introduced to quantify the recovery degree of calibration performance from super-resolved images toward high-resolution references. Experimental results demonstrate that, at a ×4 scaling factor, GC-SRNet achieves a PSNR of 35.68 dB, an SSIM of 91.55%, and a TSR of 99.15%. Field experiments further show that the calibration reprojection error is reduced from 3.2711 pixels to 0.2139 pixels for a binocular camera system with a resolution of 1280 × 960 pixels. The results indicate that GC-SRNet effectively preserves geometric fidelity while improving image quality, providing a practical solution for high-precision binocular camera calibration in vision-guided intelligent construction applications. | ||
| Key Words | ||
| binocular vision; camera calibration; construction robotics; generative adversarial networks; geometric constraints; image resolution | ||
| Address | ||
| (1) Lizhi Long, Lu Deng, Shuo Wang: College of Civil Engineering, Hunan University, Changsha, China; (2) Weiqi Mao: China Railway Major Bridge Engineering Group CO., LTD, China; (3) Lu Deng: Key Laboratory of Damage Diagnosis for Engineering Structures of Hunan Province, Hunan University, Changsha, China; (4) Lu Deng: State Key Laboratory of Bridge Safety and Resilience, Hunan University, Changsha, China; (5) Yuanfeng Duan: College of Civil Engineering and Architecture, Guangxi University, Nanning, China. | ||