随着口腔医疗领域对高精度三维重建需求的不断增长，传统多视图重建技术在处理口腔内部复杂环境时面临诸多挑战。口腔空间狭窄、采集过程抖动、牙齿表面具有高反射性等因素会导致RGB内窥镜难以均匀采样，进而引发特征匹配困难、重建结果中出现伪影噪声的问题，严重影响三维模型的精度和可靠性。针对这些问题，本文旨在提升口腔三维重建的精度、效率和鲁棒性，提出了一系列改进算法：

首先，针对口腔内部采样率不均匀导致的SuperGlue算法特征匹配质量不佳的问题，对SuperGlue算法进行了改进，提出了一种基于密度感知和置信度优化的口腔特征匹配算法。本文设计了密度感知机制来计算特征点的局部密度，用于平衡不同区域特征点的贡献；同时设计了置信度评分机制，通过综合特征描述子相似度、空间关系和密度权重，计算匹配对的置信度得分，以强化对高置信度匹配对的关注；最后设计了动态阈值调整策略，根据当前匹配对的平均密度权重和置信度得分，自适应调整匹配阈值，以满足口腔场景中采样率不均匀情况下的匹配需求。实验结果表明，改进后的算法在匹配精度上相比 SuperGlue 平均提高了约 3.65%，匹配数量增加了约 7.70%，重建后的重投影误差降低了约 3.03%，时间效率提升了约 39.99%；与其他主流算法相比，匹配精度平均提升了约 5.32%。

其次，针对高斯泼溅算法在处理口腔三维重建时无法有效应对采样率不均匀导致的伪影和噪声问题，本文对高斯泼溅算法进行了改进，提出了一种融合注意力机制的平滑口腔三维重建算法。该算法通过自注意力机制计算高斯原语之间的关联性，自适应调整原语的权重，使模型关注关键区域，抑制采样率不均匀带来的不利影响；在高斯原语优化阶段，采用三维双边滤波器，结合空间位置和特征相似性，对高斯原语进行自适应平滑处理，减少噪声同时保留重要的几何细节；在渲染阶段，应用二维双边滤波器，通过考虑像素的空间邻近性和灰度相似性，对生成的图像进行平滑处理，降低高光反射和复杂纹理区域的噪声，提升视觉质量；此外，通过平滑 L1 损失函数在小误差时提供平滑梯度以学习细节，并在大误差时保持稳健性，从而减轻异常值对模型训练的影响。实验结果显示，改进后的算法在峰值信噪比上相比原始高斯泼溅算法平均提高了约 17.89%，结构相似性指数提升了约 5.63%，感知图像相似度降低了约 18.21%，时间效率提升了约41.39%；与其他主流算法相比，重建质量平均提升了约 15.18%，建模误差平均降低了约 11.02%。

With the increasing demand for high-precision 3D reconstruction in the field of oral healthcare, traditional multi-view reconstruction techniques face numerous challenges when dealing with the complex environment inside the oral cavity. Factors such as the narrow space, motion during data acquisition, and the highly reflective surface of teeth can result in uneven sampling by RGB endoscopes, leading to difficulties in feature matching and the presence of artifacts and noise in the reconstruction results, which severely impact the accuracy and reliability of the 3D model. To address these issues, this paper aims to improve the accuracy, efficiency, and robustness of oral 3D reconstruction by proposing a series of algorithmic enhancements:

To address the issue of poor feature matching quality in the SuperGlue algorithm caused by uneven sampling rates within the oral cavity, an improved feature matching algorithm based on density awareness and confidence optimization is proposed. This study introduces a density-aware mechanism to compute the local density of feature points, balancing the contributions of feature points in different regions. Additionally, a confidence scoring mechanism is designed to calculate the confidence score of matched pairs by integrating feature descriptor similarity, spatial relationships, and density weights, thereby emphasizing high-confidence matches. A dynamic threshold adjustment strategy is also developed to adaptively adjust the matching threshold based on the average density weight and confidence score of the current matches, catering to the matching demands in oral scenarios with uneven sampling rates. Experimental results demonstrate that the improved algorithm increases matching accuracy by approximately 3.65% and the number of matches by about 7.70% compared to SuperGlue. It reduces the reprojection error of the reconstructed models by around 3.03% and improves time efficiency by approximately 39.99%. Furthermore, compared to other state-of-the-art algorithms, the matching accuracy is enhanced by an average of about 5.32%.

Secondly, to handle artifacts and noise in oral 3D reconstruction caused by uneven sampling rates—issues that the Gaussian splatting algorithm struggles to effectively address—this paper proposes an enhanced smoothing algorithm for 3D oral reconstruction that incorporates an attention mechanism. The algorithm uses a self-attention mechanism to calculate relationships among Gaussian primitives, adaptively adjusting the weights of primitives so that the model focuses on key areas and mitigates the adverse effects of uneven sampling rates. During the optimization of Gaussian primitives, a 3D bilateral filter is applied, combining spatial location and feature similarity to adaptively smooth Gaussian primitives, reducing noise while preserving significant geometric details. At the rendering stage, a 2D bilateral filter is applied to smooth the generated images by considering pixel spatial proximity and grayscale similarity, reducing noise in high-reflection and complex texture areas and enhancing visual quality. Additionally, a smooth L1 loss function is employed to provide a smooth gradient within small error ranges, aiding in detail learning, while maintaining robustness within larger error ranges to lessen the adverse effects of outliers on model training. Experimental results indicate that the improved algorithm enhances peak signal-to-noise ratio by approximately 17.89%, improves structural similarity index by about 5.63%, reduces perceptual image similarity by roughly 18.21%, and increases time efficiency by around 41.39%; compared with other mainstream algorithms, reconstruction quality improves by an average of 15.18%, and modeling error decreases by approximately 11.02%.