Deep-Learning Feature Detection for Distortion-Resistant 3D Oral Scan Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for matching three-dimensional intraoral scan data suffer from geometric distortion due to accumulated errors, particularly when using the iterative closest point (ICP) algorithm.
Innovation Solution
A method involving deep learning-based 3D feature point detection, including generating a full mouth image, detecting feature points using deep learning, determining feature points of scanned frames using a virtual frame, and re-matching frames to reconstruct a 3D oral cavity model, utilizing a device with components for matching, deep learning, virtual frame generation, and re-matching units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the iterative closest point (ICP) algorithm is used for matching point clouds, then the matching process is simple and widely applicable, but geometric distortion occurs due to accumulated errors in the reconstructed 3D model
Solution Approach 1:
The patent applies preliminary action by detecting feature points from a full mouth image before performing frame matching. This pre-detection of reliable feature points (teeth contacts, tooth centers, cusps) guides the subsequent matching process, preventing accumulated errors that would otherwise occur with standard ICP algorithms. The feature points serve as predetermined reference markers that constrain the matching to anatomically correct positions.
Solution Approach 2:
The patent introduces an intermediary element - the full mouth image with detected feature points - that mediates between the raw scan frames and the final 3D reconstruction. This intermediary provides a global reference framework that eliminates the accumulated errors inherent in direct frame-to-frame matching, while still allowing the use of simplified matching algorithms.
2Measurement precision
If deep learning-based feature point detection is applied, then matching accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent extracts only the essential feature points (teeth contacts, tooth centers, cusps) from the full mouth image using deep learning, rather than processing all points in the scan data. This selective extraction of critical anatomical landmarks provides high accuracy while limiting the computational burden to only the most important features needed for accurate matching.
Solution Approach 2:
The deep learning model serves multiple functions: it detects feature points, validates their anatomical correctness, and provides reference positions for matching. This multi-functionality consolidates several processing steps into a single comprehensive operation, improving accuracy without proportionally increasing overall system complexity.
3Reliability
If feature points are determined using a virtual frame generated from the full mouth image, then matching reliability is improved, but the processing steps and time increase
Solution Approach 1:
The virtual frame with projected feature points is generated in advance as a reference framework before performing the actual frame matching. This preliminary construction of the reference system ensures that all subsequent matching operations have a reliable, pre-established coordinate system, eliminating the need for repeated reference updates and improving overall matching reliability.
Data Source
AI summary
Disclosed is a three-dimensional oral scan data matching device including a matching unit, a deep-learning unit, a scanned frame feature determination unit, and a scan data re-matching unit. The matching unit matches a plurality of scanned frames to generate a full mouth image. The deep-learning unit performs deep-learning to detect a feature of the full mouth image. The scanned frame feature determination unit determines a feature of the plurality of scanned frames by utilizing the feature of the full mouth image. The scan data re-matching unit re-matches the plurality of scanned frames on the basis of the feature of the plurality of scanned frames, to reconstruct a three-dimensional oral model.