Eye Gaze Correction via Machine Learning Prediction

The patent extracts and isolates the eye region from the full face image, focusing computational resources only on the relevant area for gaze correction. This allows accurate gaze correction using a single webcam by concentrating processing on eye pixels rather than the entire image, resolving the contradiction between accuracy and device complexity.

The patent replaces mechanical/optical correction systems (semiransparent mirrors, stereocameras) with a computational approach using machine learning models. The neural network predicts gaze direction from eye image features, substituting physical correction mechanisms with algorithmic processing, thereby achieving accurate gaze correction with simpler hardware.

The patent performs preliminary training of the neural network model using prerecorded imagery data, but this training is done once offline. During actual video conferencing, the pre-trained model processes images in real-time without requiring additional prerecording, thus eliminating the time loss during live sessions while maintaining correction accuracy through the pre-learned model.

The patent implements a dynamic approach where the system adapts to different users and conditions. The model can be retrained or fine-tuned when needed, but operates dynamically in real-time during video calls, adjusting to varying gaze directions and lighting conditions without requiring prerecording of each session, thereby resolving the time loss issue.

The patent extracts only the eye region from the full face image and processes only this extracted portion through the neural network. This selective extraction and processing of relevant pixels significantly reduces computational energy requirements compared to processing entire face images with complex geometric modeling, while maintaining correction accuracy.

The patent uses a lightweight neural network model that can be deployed on standard consumer hardware without requiring powerful graphic accelerators. The model processes images efficiently using simplified computations rather than expensive complex geometric modeling and texture projection, reducing energy consumption while achieving practical correction accuracy.

The patent extracts and processes only the eye region independently, rather than deforming the entire face. By focusing computational processing on eye pixels separately from the rest of the face, the system achieves gaze correction without distorting global face proportions, thereby maintaining both correction accuracy and face structure integrity.

The patent applies local processing to the eye region with different transformation rules than the rest of the face. The neural network predicts gaze correction specifically for eye pixels based on local eye features, while leaving other facial regions unchanged, thus achieving gaze correction without compromising overall face proportion accuracy.