Intelligent toothbrush based on AI visual identification and oral health assessment method

Abstract

The invention provides an intelligent toothbrush based on AI visual identification and an oral health assessment method, and relates to the technical field of artificial intelligence and computer vision. The intelligent toothbrush based on AI visual identification and the oral health assessment method comprise the following steps: S1, based on a built-in multispectral camera of the intelligent toothbrush, dynamically adjusting illumination parameters of different oral regions by adopting a visible light and near-infrared bimodal synchronous acquisition algorithm and through a self-adaptive exposure compensation technology, high-resolution images and transmission data of tooth surfaces and gingival lines are obtained, and a multi-mode oral cavity image data set is generated. Illumination parameters are dynamically adjusted through a multispectral camera to obtain high-resolution bimodal data, the focus detection precision is improved in combination with cross-modal feature fusion, the problem that traditional single-modal imaging is insufficient in early-stage dental caries recognition is solved, and the accuracy of early-stage dental caries recognition is improved based on laser radar point cloud complementation and a three-dimensional surface reconstruction technology. And mapping the two-dimensional focus label into a dynamic three-dimensional plaque distribution model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and computer vision technology, and specifically to a smart toothbrush and oral health assessment method based on AI visual recognition. Background Art

[0002] The field of artificial intelligence and computer vision technology focuses on simulating human visual perception and cognitive abilities through algorithms, and combines machine learning, deep learning and other methods to process images, videos and multi-dimensional sensor data. In the field of medical and health equipment, its core technologies include: intelligent analysis of medical images (such as lesion localization and tissue segmentation), multimodal data fusion (visible light, near-infrared, lidar, etc.), three-dimensional reconstruction and dynamic modeling, and edge computing deployment optimization. Among them, the smart toothbrush and oral health assessment method based on AI visual recognition refers to an oral care system that integrates micro-optical sensors and artificial intelligence algorithms. Its purpose is to achieve quantitative evaluation of plaque distribution, early caries warning, personalized brushing guidance and other functions through real-time collection and analysis of tooth surface images, brushing behavior data and three-dimensional oral models, thereby improving users' oral cleaning efficiency, reducing the risk of periodontal disease, and providing data support for long-term health management.

[0003] Existing technologies rely on single-modality imaging data (such as visible light or near-infrared only), which is easily affected by uneven lighting and saliva interference in complex oral environments, resulting in a high rate of missed detection of subtle lesions such as early enamel demineralization and interproximal plaque. Two-dimensional image analysis has difficulty restoring the three-dimensional curvature characteristics of teeth, and is unable to quantify the true accumulation of plaque in hidden areas such as crevices and pits, resulting in deviations in the evaluation of cleaning effects. Traditional lesion detection algorithms only locate and classify isolated targets, lack systematic modeling of the tooth-gum topological relationship, and are unable to predict the risk of local plaque accumulation on the spread of gingival inflammation, resulting in delayed health warnings. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an intelligent toothbrush and oral health assessment method based on AI visual recognition, which solves the problem that the existing technology relies on single-modality imaging data (such as only visible light or near-infrared), is easily affected by uneven lighting and saliva interference in complex oral environments, resulting in a high rate of missed detection of subtle lesions such as early enamel demineralization and proximal surface plaque.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an oral health assessment method based on AI visual recognition, comprising the following steps:

[0006] S1: Based on the built-in multispectral camera of the smart toothbrush, it adopts a dual-modal synchronous acquisition algorithm of visible light and near-infrared light, and dynamically adjusts the lighting parameters of different oral areas through adaptive exposure compensation technology to obtain high-resolution images and transmission data of the tooth surface and gum line, generating a multimodal oral imaging dataset;

[0007] S2: Based on a multimodal oral imaging dataset, an improved version of YOLOv7-T iny is used, along with a cross-modal feature fusion module, to locate and classify lesions such as caries, calculus, and gingival redness and swelling, generating a heat map of oral lesions.

[0008] S3: Based on the oral lesion heat map, the LiDAR point cloud completion algorithm and 3D surface reconstruction technology are used to map the 2D lesion annotations to the 3D tooth surface model, calculate the plaque coverage area and thickness distribution, and generate a dynamic 3D plaque distribution model;

[0009] S4: Based on a dynamic three-dimensional plaque distribution model, a graph convolutional neural network is used to model the tooth-gum topological relationship. This method analyzes the impact of plaque accumulation at specific tooth sites on the health of adjacent gums, predicts potential inflammation risk areas, and generates an oral health relationship map.

[0010] S5: Based on the oral health relationship map and the timestamp information of S1, a transfer learning framework is used, and through the fusion of brushing pressure sensor data, the user's brushing trajectory is optimized, personalized action correction suggestions are generated, and a set of brushing behavior optimization parameters is generated;

[0011] S6: Based on the brushing behavior optimization parameter set, it adopts edge computing lightweight deployment technology and transmits the parameters to the smart toothbrush MCU in real time through the low-power Bluetooth communication protocol, dynamically adjusts the vibration frequency and motion trajectory, and simultaneously generates a 3D visual health report on the APP side, generating a real-time oral health feedback system.

[0012] Preferably, generating a multimodal oral image dataset based on S1 includes the following steps:

[0013] S101: Based on the multispectral camera array on the smart toothbrush head, it uses a dual-channel synchronous triggering algorithm for visible light and near-infrared light, and uses dynamic area exposure control technology to capture tooth surface texture and dentin transmission images to generate original dual-modal oral image pairs;

[0014] S102: Based on the original bimodal oral image pair, the two modal images are aligned using an affine transformation registration algorithm. The lesion edge features are enhanced through hybrid Gaussian-Laplacian pyramid fusion to eliminate motion blur and uneven illumination interference, and generate a registered enhanced multimodal oral image dataset.

[0015] Preferably, generating an oral lesion heat map based on S2 comprises the following steps:

[0016] S201: Based on the registration-enhanced multimodal oral image dataset, a two-stream convolutional neural network is used to extract visible light texture features and near-infrared transmission features respectively. The channel attention weighted module is then used to filter key feature maps to generate multimodal lesion feature maps.

[0017] S202: Based on the multimodal lesion feature map, the improved YOLOv7-T iny target detection network is used, and the bimodal features are fused through the Transformer cross-attention mechanism to calculate the lesion bounding box and category confidence to generate an oral lesion heat map.

[0018] Preferably, generating a dynamic three-dimensional plaque distribution model based on S3 includes the following steps:

[0019] S301: Based on the oral lesion heat map, the LiDAR point cloud sparse completion algorithm is used to map the 2D lesion coordinates to the 3D point cloud space through iterative closest point registration technology, complete the point cloud data of the adjacent surface blind area, and generate a 3D lesion point cloud completion model;

[0020] S302: Based on the 3D lesion point cloud completion model, a Poisson surface reconstruction algorithm is used to generate a continuous tooth surface topology structure. The thickness gradient of the plaque coverage area is annotated through the U-Net semantic segmentation network to generate a dynamic 3D plaque distribution model.

[0021] Preferably, generating an oral health relationship map based on S4 includes the following steps:

[0022] S401: Based on a dynamic three-dimensional plaque distribution model, a heterogeneous graph construction algorithm is used to model each tooth as a graph node, and the gingival sulcus depth and saliva pH value as edge attributes. A tooth-gum interaction topological network is established to generate a heterogeneous oral health graph structure.

[0023] S402: Based on the oral health heterogeneous graph structure, a graph convolutional neural network combined with a graph attention mechanism is used to calculate the probability of plaque accumulation at a specific tooth site spreading to adjacent gingival inflammation, output high-risk areas and their impact paths, and generate an oral health relationship map.

[0024] Preferably, generating the toothbrushing behavior optimization parameter set based on S5 includes the following steps:

[0025] S501: Based on the oral health relationship graph and the timestamp information of S1, a domain adversarial neural network is used to align the user's mobile phone gyroscope data with the toothbrush motion sensor data, extract cross-device consistency features such as brushing angle and speed, and generate a cross-device brushing behavior feature vector;

[0026] S502: Based on the cross-device brushing behavior feature vector, a reinforcement learning framework is adopted, and plaque removal efficiency is used as the reward function to dynamically optimize the brushing path and generate a brushing behavior optimization parameter set.

[0027] Preferably, generating a real-time oral health feedback system based on S6 includes the following steps:

[0028] S601: Based on a parameter set optimized for brushing behavior, the ResNet-50 teacher model is compressed into a MobileNetV3 student model using knowledge distillation technology. This model is then deployed to the toothbrush-side ARM Cortex-M7 chip using the TensorRT acceleration engine to generate a lightweight edge inference model.

[0029] S602: Based on a lightweight edge inference model and using the low-power Bluetooth Mesh network protocol, the optimized parameters are transmitted to the toothbrush motor control system in real time. The Three.js engine is simultaneously called on the mobile app to render a 3D health report, generating a real-time oral health feedback system.

[0030] A smart toothbrush based on AI visual recognition, comprising the oral health assessment method based on AI visual recognition according to any one of claims 1 to 7.

[0031] The present invention provides a smart toothbrush and oral health assessment method based on AI visual recognition. It has the following beneficial effects:

[0032] The present invention uses a multispectral camera to dynamically adjust illumination parameters to obtain high-resolution bimodal data, and combines cross-modal feature fusion to improve lesion detection accuracy, thus solving the problem of insufficient early caries identification in traditional single-modal imaging. Based on LiDAR point cloud completion and three-dimensional surface reconstruction technology, the two-dimensional lesion annotation is mapped into a dynamic three-dimensional plaque distribution model, breaking through the blind spots of plane analysis in the detection of adjacent surfaces and pits and fissures, quantifying the plaque thickness and coverage area, and realizing a visual assessment of the cleaning effect of hidden areas. The tooth-gum topological relationship modeling and graph convolutional neural network are introduced to analyze the systemic impact of plaque accumulation on gum health, predict potential inflammatory risk paths, and make up for the lack of correlation analysis in traditional isolated lesion detection. The timestamp information and pressure sensor data are integrated to optimize the brushing trajectory, and the equipment parameters are adjusted in real time through lightweight deployment of edge computing. A 3D visualization report is generated synchronously, forming a closed-loop feedback from data collection to behavior correction, thereby improving the timeliness and accuracy of personalized intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0034] Figure 2 This is a schematic diagram of the refinement of S1 of the present invention;

[0035] Figure 3 This is a schematic diagram of the refinement of S2 of the present invention;

[0036] Figure 4 This is a schematic diagram of the refinement of S3 of the present invention;

[0037] Figure 5 This is a schematic diagram of the refinement of S4 of the present invention;

[0038] Figure 6 This is a schematic diagram of the refinement of S5 of the present invention;

[0039] Figure 7 This is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] Example:

[0042] like Figure 1-7 As shown, an embodiment of the present invention provides a cavity health assessment method based on AI visual recognition, comprising the following steps:

[0043] S1: Based on the built-in multispectral camera of the smart toothbrush, it adopts a dual-modal synchronous acquisition algorithm of visible light and near-infrared light, and dynamically adjusts the lighting parameters of different oral areas through adaptive exposure compensation technology to obtain high-resolution images and transmission data of the tooth surface and gum line, generating a multimodal oral imaging dataset;

[0044] S2: Based on a multimodal oral imaging dataset, an improved version of YOLOv7-Tiny is used, along with a cross-modal feature fusion module, to locate and classify lesions such as caries, calculus, and gingival redness and swelling, generating a heat map of oral lesions.

[0045] S3: Based on the oral lesion heat map, the LiDAR point cloud completion algorithm and 3D surface reconstruction technology are used to map the 2D lesion annotations to the 3D tooth surface model, calculate the plaque coverage area and thickness distribution, and generate a dynamic 3D plaque distribution model;

[0046] S4: Based on a dynamic three-dimensional plaque distribution model, a graph convolutional neural network is used to model the tooth-gum topological relationship. This method analyzes the impact of plaque accumulation at specific tooth sites on the health of adjacent gums, predicts potential inflammation risk areas, and generates an oral health relationship map.

[0047] S5: Based on the oral health relationship map and the timestamp information of S1, a transfer learning framework is used, and through the fusion of brushing pressure sensor data, the user's brushing trajectory is optimized, personalized action correction suggestions are generated, and a set of brushing behavior optimization parameters is generated;

[0048] S6: Based on the brushing behavior optimization parameter set, it adopts edge computing lightweight deployment technology and transmits the parameters to the smart toothbrush MCU in real time through the low-power Bluetooth communication protocol, dynamically adjusts the vibration frequency and motion trajectory, and simultaneously generates a 3D visual health report on the APP side, generating a real-time oral health feedback system.

[0049] Generating a multimodal oral imaging dataset based on S1 includes the following steps:

[0050] S101: Based on the multispectral camera array on the smart toothbrush head, it uses a dual-channel synchronous triggering algorithm for visible light and near-infrared light, and uses dynamic area exposure control technology to capture tooth surface texture and dentin transmission images to generate original dual-modal oral image pairs;

[0051] A miniature multispectral camera array (comprising visible light CMOS and near-infrared sensors) is embedded in the head of the smart toothbrush. Using dual-channel time-sharing multiplexing, it alternately triggers two imaging modalities with 10ms-level time synchronization accuracy. Dynamic area exposure control automatically reduces exposure intensity for highly reflective areas (such as enamel) and increases near-infrared light source power to 850mW for low-transmittance areas (such as proximal surfaces), based on the difference in tooth surface reflectivity (visible light) and dentin transmittance (near-infrared).

[0052] S102: Based on the original bimodal oral image pair, the two modal images are aligned using an affine transformation registration algorithm. The lesion edge features are enhanced through hybrid Gaussian-Laplacian pyramid fusion to eliminate motion blur and uneven illumination interference, and generate a registered enhanced multimodal oral image dataset.

[0053] To address the spatial offset problem between visible and near-infrared images, an affine transformation registration algorithm was used to calculate the rotation and translation matrix between the two modalities (error ≤ 0.5 pixels). Furthermore, a hybrid Gaussian-Laplacian pyramid fusion was used to layer-fuse visible light high-frequency textures (such as enamel cracks) with near-infrared low-frequency transmission features (such as dentin mineralization). Guided filtering was used to eliminate motion blur artifacts.

[0054] Generating oral lesion heatmap based on S2 includes the following steps;

[0055] S201: Based on the registration-enhanced multimodal oral image dataset, a two-stream convolutional neural network is used to extract visible light texture features and near-infrared transmission features respectively. The channel attention weighted module is then used to filter key feature maps to generate multimodal lesion feature maps.

[0056] A two-stream convolutional neural network is constructed, in which the visible light branch uses depthwise separable convolution to extract surface texture (convolution kernel 3×3, stride 2), and the near-infrared branch uses void convolution (dilation rate 2) to capture deep transmission features. Feature weights are dynamically allocated through the channel attention weighting module to suppress interference areas such as saliva reflections.

[0057] S202: Based on the multimodal lesion feature map, the improved YOLOv7-T iny target detection network is used, and the bimodal features are fused through the Transformer cross-attention mechanism to calculate the lesion bounding box and category confidence to generate an oral lesion heat map.

[0058] Based on YOLOv7-T iny, the Transformer cross-attention mechanism is introduced. The visible light feature map is used as the query and the near-infrared feature map is used as the key-value. The cross-modal correlation weight is calculated. The detection of small targets (such as early white spots of dental caries) is optimized through a dynamic label assignment strategy, and the Gaussian kernel density estimation is used to generate the heat map confidence distribution.

[0059] Generating a dynamic three-dimensional plaque distribution model based on S3 includes the following steps:

[0060] S301: Based on the oral lesion heat map, the LiDAR point cloud sparse completion algorithm is used to map the 2D lesion coordinates to the 3D point cloud space through iterative closest point registration technology, complete the point cloud data of the adjacent surface blind area, and generate a 3D lesion point cloud completion model;

[0061] The improved PointNet++ is used to process lidar point cloud data. The local geometric structure features are learned through k-nearest neighbor graph convolution. Combined with the semantic prior information of the two-dimensional heat map, the generative adversarial network is used to complete the point clouds of blind spots such as adjacent surfaces and pits and grooves (the completion accuracy reaches 98.5%).

[0062] S302: Based on the 3D lesion point cloud completion model, a Poisson surface reconstruction algorithm is used to generate a continuous tooth surface topology structure. The thickness gradient of the plaque coverage area is annotated through the U-Net semantic segmentation network to generate a dynamic 3D plaque distribution model.

[0063] A continuous tooth surface mesh model (number of triangular facets ≤ 500,000) was generated based on Poisson surface reconstruction, and the plaque thickness gradient was analyzed using the U-Net++ semantic segmentation network: multispectral images and point cloud curvature data were input, and a thickness distribution map (0-1 mm classification, resolution 0.05 mm) was output.

[0064] Generating an oral health relationship map based on S4 includes the following steps:

[0065] S401: Based on a dynamic three-dimensional plaque distribution model, a heterogeneous graph construction algorithm is used to model each tooth as a graph node, and the gingival sulcus depth and saliva pH value as edge attributes. A tooth-gum interaction topological network is established to generate a heterogeneous oral health graph structure.

[0066] Each tooth is modeled as a graph node, with attributes including plaque thickness, interproximal contact angle, and root morphology coefficient; gingival margin attributes include groove depth (0-3mm grading) and pH value (5.5-7.0 dynamic monitoring data). The mechanical effects between teeth and the microbial migration path are described through a dynamic adjacency matrix.

[0067] S402: Based on the oral health heterogeneous graph structure, a graph convolutional neural network combined with a graph attention mechanism is used to calculate the probability of plaque accumulation at a specific tooth site spreading to adjacent gingival inflammation, output high-risk areas and their impact paths, and generate an oral health relationship map.

[0068] Based on GCN, a graph attention mechanism is introduced to calculate the attention coefficient of the plaque accumulation node to the adjacent gingival node (weight range 0-1), and the probability of inflammation spread in the next 48 hours is predicted through the time series propagation model, and high-risk paths are output (such as the probability of premolar plaque causing adjacent gingival recession ≥75%).

[0069] Generating an optimized parameter set for tooth brushing behavior based on S5 includes the following steps:

[0070] S501: Based on the oral health relationship graph and the timestamp information of S1, a domain adversarial neural network is used to align the user's mobile phone gyroscope data with the toothbrush motion sensor data, extract cross-device consistency features such as brushing angle and speed, and generate a cross-device brushing behavior feature vector;

[0071] A domain adversarial neural network is used to align the data of the mobile phone gyroscope (sampling rate 100Hz) and the toothbrush IMU sensor (sampling rate 500Hz). The maximum mean difference loss function is used to eliminate the baseline drift between the devices, and consistency features such as brushing angle (±30°) and path curvature (0-1.2) are extracted.

[0072] S502: Based on the cross-device brushing behavior feature vector, a reinforcement learning framework is adopted, and plaque removal efficiency is used as the reward function to dynamically optimize the brushing path and generate a brushing behavior optimization parameter set.

[0073] A PPO reinforcement learning framework was constructed. The state space included plaque distribution and gum risk index. The action space defined eight brush head motion modes (such as the modified trajectory of the Bass method). The reward function integrated the plaque removal rate (weight 0.6) and the gum damage value (weight 0.4). The Pareto optimal solution was generated through Monte Carlo tree search.

[0074] Generating a real-time oral health feedback system based on S6 includes the following steps;

[0075] S601: Based on a parameter set optimized for brushing behavior, the ResNet-50 teacher model is compressed into a MobileNetV3 student model using knowledge distillation technology. This model is then deployed to the toothbrush-side ARM Cortex-M7 chip using the TensorRT acceleration engine to generate a lightweight edge inference model.

[0076] Using progressive knowledge distillation, the hidden layer attention map of the ResNet-50 teacher model is migrated to MobileNetV3, quantized to INT8 precision (accuracy loss ≤ 2%) through the TensorRT engine, and achieved 10ms inference latency on the ARM Cortex-M7 chip (power consumption ≤ 0.3W).

[0077] S602: Based on a lightweight edge inference model and using the low-power Bluetooth Mesh network protocol, the optimized parameters are transmitted to the toothbrush motor control system in real time. The Three.js engine is simultaneously called on the mobile app to render a 3D health report, generating a real-time oral health feedback system.

[0078] Based on the low-power Bluetooth Mesh protocol (transmission interval 20ms), the optimized parameters are broadcast to the toothbrush motor (response time ≤ 5ms), and the Three.js engine is synchronously called to render a 3D oral health report: the plaque distribution is colored in a red-yellow-green gradient, the risk path is dynamically marked with particle flow, and 720° viewing angle interaction is supported.

[0079] A smart toothbrush based on AI visual recognition, comprising the oral health assessment method based on AI visual recognition according to any one of claims 1 to 7.

[0080] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Oral health assessment method based on AI visual recognition, characterized by: The following steps are involved: S1: Based on the built-in multispectral camera of the smart toothbrush, it adopts a dual-modal synchronous acquisition algorithm of visible light and near-infrared light, and dynamically adjusts the lighting parameters of different oral areas through adaptive exposure compensation technology to obtain high-resolution images and transmission data of the tooth surface and gum line, generating a multimodal oral imaging dataset; S2: Based on a multimodal oral imaging dataset, an improved version of YOLOv7-Tiny is used, along with a cross-modal feature fusion module, to locate and classify lesions such as caries, calculus, and gingival redness and swelling, generating a heat map of oral lesions. S3: Based on the oral lesion heat map, the LiDAR point cloud completion algorithm and 3D surface reconstruction technology are used to map the 2D lesion annotations to the 3D tooth surface model, calculate the plaque coverage area and thickness distribution, and generate a dynamic 3D plaque distribution model; S4: Based on a dynamic three-dimensional plaque distribution model, a graph convolutional neural network is used to model the tooth-gum topological relationship. This method analyzes the impact of plaque accumulation at specific tooth sites on the health of adjacent gums, predicts potential inflammation risk areas, and generates an oral health relationship map. S5: Based on the oral health relationship map and the timestamp information of S1, a transfer learning framework is used, and through the fusion of brushing pressure sensor data, the user's brushing trajectory is optimized, personalized action correction suggestions are generated, and a set of brushing behavior optimization parameters is generated; S6: Based on the brushing behavior optimization parameter set, it adopts edge computing lightweight deployment technology and transmits the parameters to the smart toothbrush MCU in real time through the low-power Bluetooth communication protocol, dynamically adjusts the vibration frequency and motion trajectory, and simultaneously generates a 3D visual health report on the APP side, generating a real-time oral health feedback system.

2. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating a multimodal oral imaging dataset based on S1 includes the following steps: S101: Based on the multispectral camera array on the smart toothbrush head, it uses a dual-channel synchronous triggering algorithm for visible light and near-infrared light, and uses dynamic area exposure control technology to capture tooth surface texture and dentin transmission images to generate original dual-modal oral image pairs; S102: Based on the original bimodal oral image pair, the two modal images are aligned using an affine transformation registration algorithm. The lesion edge features are enhanced through hybrid Gaussian-Laplacian pyramid fusion to eliminate motion blur and uneven illumination interference, and generate a registered enhanced multimodal oral image dataset.

3. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating oral lesion heatmap based on S2 includes the following steps; S201: Based on the registration-enhanced multimodal oral image dataset, a two-stream convolutional neural network is used to extract visible light texture features and near-infrared transmission features respectively. The channel attention weighted module is then used to filter key feature maps to generate multimodal lesion feature maps. S202: Based on the multimodal lesion feature map, the improved YOLOv7-T iny target detection network is used, and the bimodal features are fused through the Transformer cross-attention mechanism to calculate the lesion bounding box and category confidence to generate an oral lesion heat map.

4. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating a dynamic three-dimensional plaque distribution model based on S3 includes the following steps: S301: Based on the oral lesion heat map, the LiDAR point cloud sparse completion algorithm is used to map the 2D lesion coordinates to the 3D point cloud space through iterative closest point registration technology, complete the point cloud data of the adjacent surface blind area, and generate a 3D lesion point cloud completion model; S302: Based on the 3D lesion point cloud completion model, a Poisson surface reconstruction algorithm is used to generate a continuous tooth surface topology structure. The thickness gradient of the plaque coverage area is annotated through the U-Net semantic segmentation network to generate a dynamic 3D plaque distribution model.

5. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating an oral health relationship map based on S4 includes the following steps: S401: Based on a dynamic three-dimensional plaque distribution model, a heterogeneous graph construction algorithm is used to model each tooth as a graph node, and the gingival sulcus depth and saliva pH value as edge attributes. A tooth-gum interaction topological network is established to generate a heterogeneous oral health graph structure. S402: Based on the oral health heterogeneous graph structure, a graph convolutional neural network combined with a graph attention mechanism is used to calculate the probability of plaque accumulation at a specific tooth site spreading to adjacent gingival inflammation, output high-risk areas and their impact paths, and generate an oral health relationship map.

6. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating an optimized parameter set for tooth brushing behavior based on S5 includes the following steps: S501: Based on the oral health relationship graph and the timestamp information of S1, a domain adversarial neural network is used to align the user's mobile phone gyroscope data with the toothbrush motion sensor data, extract cross-device consistency features such as brushing angle and speed, and generate a cross-device brushing behavior feature vector; S502: Based on the cross-device brushing behavior feature vector, a reinforcement learning framework is adopted, and plaque removal efficiency is used as the reward function to dynamically optimize the brushing path and generate a brushing behavior optimization parameter set.

7. The oral health assessment method based on AI visual recognition according to claim 1, characterized in that: Generating a real-time oral health feedback system based on S6 includes the following steps; S601: Based on a parameter set optimized for brushing behavior, the ResNet-50 teacher model is compressed into a MobileNetV3 student model using knowledge distillation technology. This model is then deployed to the toothbrush-side ARM Cortex-M7 chip using the TensorRT acceleration engine to generate a lightweight edge inference model. S602: Based on a lightweight edge inference model and using the low-power Bluetooth Mesh network protocol, the optimized parameters are transmitted to the toothbrush motor control system in real time. The Three.js engine is simultaneously called on the mobile app to render a 3D health report, generating a real-time oral health feedback system.

8. The smart toothbrush based on AI visual recognition is characterized by: Including the oral health assessment method based on AI visual recognition as described in any one of claims 1-7.

Images