@article{TIAN2026116128,
title = {Specular-suppressed three-dimensional Gaussian splatting for transparent object reconstruction},
journal = {Engineering Applications of Artificial Intelligence},
volume = {182},
pages = {116128},
year = {2026},
issn = {0952-1976},
doi = {https://doi.org/10.1016/j.engappai.2026.116128},
url = {https://www.sciencedirect.com/science/article/pii/S0952197626024127},
author = {Chenyu Tian and Wentao Hu and Long Wen and Huafeng Ding},
keywords = {Three-dimensional reconstruction, Transparent objects, Gaussian splatting, Virtual three-dimensional models, Artificial intelligence},
abstract = {Accurate three-dimensional reconstruction of transparent objects remains a critical challenge in engineering applications such as industrial inspection, robotic perception, and automated manufacturing systems, due to complex light transport effects including refraction, reflection, and specular highlights. These effects violate photometric consistency assumptions and significantly limit the performance of existing neural rendering and three-dimensional Gaussian splatting methods. This paper presents an artificial intelligence-based three-dimensional Gaussian splatting framework for transparent object reconstruction with explicit specular suppression. From the perspective of artificial intelligence, the main contribution lies in: (1) a specular-aware decomposition module is first designed to separate diffuse and specular signals while estimating pixel-wise specular reliability; (2) a specular Gaussian refinement strategy is introduced to selectively update geometry-related Gaussian parameters through confidence-guided optimization; (3) a geometry-aware regularization module is constructed to incorporate surface constraints. From the perspective of engineering application, the proposed method enables more accurate and stable reconstruction of transparent objects under complex illumination conditions, providing a potentially more reliable geometric representation for downstream inspection and perception tasks. Extensive experiments on synthetic and real-world datasets demonstrate that the proposed method achieves the best overall average performance among the compared methods, particularly in geometric reconstruction accuracy, while remaining competitive in rendering quality. These results indicate that the proposed framework provides an effective solution for transparent object reconstruction under specular interference and shows strong potential for practical engineering applications.}
}