Neural Network 3D Reconstruction for Ambiguous Pattern Correspondence

The patent replaces traditional geometric calculation methods with a neural network-based system to determine correspondence between image pixels and projection pattern elements. The neural network processes distorted projection patterns and outputs accurate correspondence relationships, substituting complex geometric algorithms with a trained neural network model that achieves superior precision in handling non-coded projection elements.

The patent uses simulated training data that copies real-world projection patterns and their corresponding correspondence relationships to train the neural network. By training on synthetic data representing various objects and projection scenarios, the neural network learns to accurately match image pixels with projection pattern elements without requiring complex real-time calculations during actual scanning.

The patent divides the correspondence determination task into multiple stages by using multiple neural networks in a cascaded manner. Each neural network processes specific aspects of the projection pattern distortion and contributes to the final correspondence mapping. This segmentation allows each network to specialize in specific features or distortion types, improving overall resolution and completeness while distributing computational load.

The patent performs preliminary training of neural networks on simulated data before actual scanning operations. The neural networks are pre-trained to recognize projection pattern distortions and determine correspondences, so that during actual use, they can quickly process real images without requiring extensive computation time for learning or adaptation.