随着口腔疾病患病率的增高，人们逐渐意识到定期口腔监测的重要性。目前大部分家庭需要前往专业口腔机构进行定期体检，耗费精力和成本。针对这一问题，需要构建一款低成本、高效率、易用性良好的家庭式智慧口腔健康服务系统来满足人们的日常健康自检需求。系统以口腔内窥镜作为硬件支撑，小程序为载体，智能拼接功能算法和智能检测功能算法为关键研究内容，并结合图像采集、结果分析与管理、知识传播、在线问诊、商品推广及消息提醒等业务功能，为不同年龄段、不同经济状况、不同文化程度的家庭用户提供更智能便捷的健康监护服务。

       在智能拼接功能算法中，针对现有拼接算法存在可匹配点对稀疏，内点误匹配，单应性矩阵鲁棒性低，大视差等问题，本文提出基于改进SIFT的全景颌面牙齿拼接算法对颌面牙齿图像进行拼接。在图像预处理部分，使用同态滤波器对感兴趣的特征进行增强，并有效去除信号中的干扰噪声，提升可匹配关键点的数目。在特征配准部分，提出基于交叉过滤的方邻域约束提纯内点对，并优化单应性矩阵，提升鲁棒性。在图像融合部分，使用基于动态规划的最佳缝合线，并结合拉普拉斯金字塔算法进行信息融合，有效缓解因相机运动视差引起的鬼影和拼接不平滑等问题。

       在智能检测功能算法中，针对现有算法存在细粒度特征提取难，语义表征能力不强，私人数据集偏见等问题，本文提出一种基于弱监督的牙齿病灶语义分割算法实现病灶检测。在特征提取部分，使用CAM掩码擦除迫使网络将更多注意力用于特征提取和语义表征上，提升网络捕获更全面特征的能力，且对数据集偏见具有鲁棒性。在生成伪标签过程中，提出注意力补偿机制和CRF后处理操作，避免了前景区域激活不足和背景区域激活过度问题，提高了伪标签的置信度。在分割训练过程中，设计能量损失函数与交叉熵损失函数联合使用，同时考虑有标签和无标签区域的利用，从而提升网络分割性能。

       系统应用全景颌面牙齿拼接算法完成多张颌面局部图像的智能拼接，生成一张完整的口腔颌面全景图，方便用户直观查看口腔整体健康状态，且便于医患在线沟通并准确了解病灶牙齿的牙位信息。应用牙齿病灶语义分割算法完成牙齿病灶区域的智能检测，为疾病的辅助诊断提供客观的医学图像分析依据。此外，结合在线问诊、知识传播、消息提醒和商品推广等智慧化服务，推动家庭口腔医疗模式从“以疾病诊疗为主”到“以预防为主”转变，提高国民口腔保健意识。

      With the increasing prevalence of oral diseases, people gradually realize the importance of regular oral monitoring. At present, most families need to go to professional oral institutions for regular examination, which consumes energy and cost. In response to this problem, a low-cost, high-efficiency and easy-to-use home-based intelligent oral health service system needs to be built to meet people's daily health self-examination needs. The system takes the oral endoscope as the hardware support, the small program as the carrier, the intelligent splicing function algorithm and the intelligent detection function algorithm as the key research contents, and combines the business functions of image acquisition, result analysis and management, knowledge dissemination, online consultation, commodity promotion and message reminder to provide more intelligent and convenient health monitoring services for home users of different ages, different economic conditions and different educational levels.

      In the intelligent splicing function algorithm, aiming at the problems of sparse matching point pairs, mismatching internal points, low robustness of homography matrix and large parallax in the existing splicing algorithm, this paper proposed an improved SIFT panoramic maxillofacial tooth splicing algorithm to splice maxillofacial tooth images. In the image preprocessing part, the homomorphic filter is used to enhance the features of interest, effectively remove the interference noise in the signal, and increase the number of matching key points. In the feature registration part, a square neighborhood constraint based on cross filtering is proposed to purify the interior point pairs, and the homography matrix is optimized to improve the robustness. In the image fusion part, the optimal stitching line based on dynamic programming is used and the Laplacian pyramid algorithm is combined for information fusion, which effectively alleviates the problems of ghosting and unsmooth stitching caused by camera motion parallax.

      In the intelligent detection function algorithm, aiming at the problems of fine-grained feature extraction, weak semantic representation ability and private dataset bias in the existing algorithms, this paper proposes a semantic segmentation algorithm for dental lesions based on weak supervision to realize lesion detection. In the feature extraction part, using CAM mask erasure forces the network to devote more attention to feature extraction and semantic representation, which improves the network's ability to capture more comprehensive features and is robust to dataset bias. In the process of generating pseudo-labels, an attention compensation mechanism and CRF post-processing operation are proposed to avoid the problem of under-activation of foreground regions and over-activation of background regions, and improve the confidence of pseudo-labels. In the segmentation training process, the energy loss function and the cross entropy loss function are designed to be used in combination, and the utilization of labeled and unlabeled regions is considered to improve the network segmentation performance.

      The system applies the panoramic maxillofacial tooth splicing algorithm to complete the intelligent splicing of multiple maxillofacial local images, and generates a complete maxillofacial tooth panoramic image, which can visually view the overall state of oral health, and facilitate online communication between doctors and patients and accurately understand the tooth position information of the lesion teeth. The semantic segmentation algorithm of dental lesions is applied to complete the intelligent detection of dental lesions, which provides objective medical image analysis basis for the auxiliary diagnosis of diseases. In addition, combined with intelligent services such as online consultation, knowledge dissemination, message reminding and product promotion, it promotes the transformation of the family oral health care model from 'disease diagnosis and treatment-oriented' to 'prevention-oriented' and improves the national oral health awareness.