Autofocus Eye Selection via Depth Estimation

The patent transitions from 2D image plane detection to 3D spatial reasoning by estimating depth distances from the camera to each eye. The neural network model predicts distance values for multiple eye candidates, enabling the system to select the closest eye based on inferred depth information rather than relying solely on 2D image coordinates and facial orientation.

The patent replaces traditional mechanical/optical autofocus detection mechanisms with an AI-based neural network approach. Instead of using phase detection sensors or contrast-based methods that struggle with angled subjects, the system uses deep learning models trained on labeled data to directly predict eye positions and distances, substituting physical detection methods with computational intelligence.

The patent applies preliminary action by pre-training neural network models on large datasets of labeled images before deployment. The models are trained offline to learn the complex mappings between image features and eye position/distance relationships, so that during actual autofocus operation, the pre-trained networks can quickly infer distances without requiring complex real-time computations.

The patent uses copying by creating synthetic training data through image processing and labeling. Ground truth distance maps are generated by processing labeled images through geometric transformations, creating copies of real scenarios with known answers. This allows the neural network to be trained on extensive synthetic data that mirrors real-world conditions without requiring equally extensive real-world annotated datasets.