Spatial Transformer Network for Face Recognition Data Efficiency

The patent applies preliminary action by pre-transforming input images into a standardized structure using a spatial transformer network before face recognition. The STN model pre-aligns facial features (eyes, nose, mouth) to canonical positions, creating a normalized representation that enables faster and more accurate recognition without requiring extensive training data. This preprocessing step prepares the data in advance, allowing the recognition system to operate efficiently on standardized inputs.

The patent applies parameter changes by transforming the input image parameters (position, scale, orientation) into a standardized canonical form using the spatial transformer network. By changing the parameter space representation of facial features to a normalized coordinate system, the system achieves high recognition accuracy with fewer training samples. The STN learns to adjust transformation parameters (translation, rotation, scaling) to align faces consistently, enabling efficient learning from limited data.

The patent applies universality by designing a spatial transformer network that learns generalizable facial feature transformations applicable across different target objects. The STN model identifies and transforms canonical facial landmarks (eyes, nose, mouth) that are universal across various objects, enabling the system to recognize faces of different target objects reliably. The learned transformation parameters and feature extraction methodology transfer across object types, providing both reliability for trained objects and adaptability for new objects.