Teeth Occlusion Detection Using Depth Maps and Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current dental alignment technologies face challenges in accurately assessing teeth alignment and generating effective treatment plans due to cumbersome calibration processes and high error rates in depth map algorithms, particularly for monocular depth maps and feature-based methods.
Innovation Solution
A method utilizing a U-Net neural network for depth map generation and a Go/No-Go classification system, which includes pre-processing for noise reduction, segmentation, and classification using convolutional neural networks to determine teeth alignment within predefined parameters, eliminating the need for explicit calibration of camera matrices and providing a 3D printable aligner structure for teeth alignment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If monocular depth map algorithms are used for teeth alignment assessment, then the process is simplified and does not require camera calibration, but the error rate increases and measurement precision deteriorates
Solution Approach 1:
The patent introduces a deep learning-based segmentation model as an intermediary between the input image and depth map generation. This model pre-processes the image to identify and segment teeth structures, providing refined input to the depth estimation algorithm. This intermediary step filters out irrelevant information and focuses the depth estimation on relevant dental structures, thereby improving precision without requiring camera calibration
Solution Approach 2:
The patent transitions from traditional 2D image processing to 3D spatial reasoning by generating depth maps that add the Z-dimension. The system creates pseudo-3D representations of teeth from 2D images, enabling volumetric analysis and more accurate alignment assessment. This dimensional enhancement allows the system to achieve medical-grade precision comparable to calibrated systems while maintaining the simplicity of monocular imaging
2Reliability
If feature-based depth map algorithms (such as MeshroomTMM using SURF or SIFT) are used, then key point matching is performed, but the result is not dense and quality is low
Solution Approach 1:
The patent replaces traditional feature-based mechanical matching methods (SURF, SIFT) with a deep learning-based end-to-end depth estimation system. Instead of manually detecting and matching key points through geometric algorithms, the system uses convolutional neural networks to directly predict depth values from image features. This substitution enables dense pixel-level depth estimation rather than sparse point-cloud-based estimation, significantly improving depth map quality and completeness
Solution Approach 2:
The patent changes the fundamental parameters of depth estimation by transitioning from discrete key point coordinates to continuous pixel-level depth values. The deep learning model outputs a dense depth map where every pixel has an associated depth value, rather than relying on sparse 3D point clouds from feature matching. This parameter transformation enables comprehensive surface reconstruction and improves overall depth map quality
3Measurement precision
If chessboard calibration is performed to obtain accurate intrinsic and extrinsic matrices, then measurement precision is improved, but the device complexity and time required increase
Solution Approach 1:
The patent extracts and removes the calibration step entirely from the imaging pipeline. Instead of requiring chessboard calibration to obtain intrinsic and extrinsic camera matrices, the system uses a deep learning model that is trained to be invariant to camera parameters. The neural network learns to estimate depth directly from image content without needing explicit camera calibration data, thereby eliminating the complex calibration process while maintaining measurement accuracy
Solution Approach 2:
The patent creates a virtual copy of the calibration process through deep learning. Rather than performing physical calibration with chessboards, the system uses a pre-trained neural network that has learned camera characteristics from大量 calibrated training data. This virtual calibration copy allows the system to operate as if calibrated without requiring actual calibration hardware or procedures, reducing device complexity while preserving measurement precision
4Measurement precision
If LSD-SLAM algorithm is used for depth optimization, then semi-dense depth estimation is achieved, but the algorithm is not fully parallel and processing speed decreases
Solution Approach 1:
The patent replaces the iterative optimization-based LSD-SLAM algorithm with a deep learning-based direct depth estimation approach. Instead of using gradient descent and Lie group optimization, the system uses a convolutional neural network that directly predicts depth in a single forward pass. This substitution maintains semi-dense depth estimation quality while enabling full parallelization on GPU hardware, dramatically improving processing speed and productivity
Data Source
AI summary
A method including: obtaining a plurality of images in which front teeth of a subject are visible; performing segmentation on selected images from the plurality of images to create a first segmentation mask and labeling each tooth in the selected images to provide a detailed segmentation map; generating a depth map of the front teeth; calculating a horizontal gradient of the depth map and a vertical moving average of a plurality of pixels of the horizontal gradient to receive depth gradients and flagging depth gradients where the vertical moving average exceeds a predefined threshold or is classified by an Artificial Neural Network or other machine learning model as abnormal; inputting the depth gradients and detailed segmentation map into a classifier to determine whether the front teeth are within predetermined parameters; and receiving a go or no-go classification from the classifier.