Image Gaze Correction Using Flow Fields and Contour Masks

The patent combines multiple sequential parameter calculations (head pose angle, head offset direction, gaze angle, eyeball position correction) into a single integrated deep learning model. The model directly maps input eye images to corrected output images in one end-to-end process, eliminating the need for separate calculation steps and preventing error accumulation while maintaining ease of implementation through a unified neural network architecture.

The patent implements continuous processing by using a deep learning model that processes the entire gaze correction task in a single forward pass. Instead of executing multiple discrete calculation steps that require intermediate storage and repeated data processing, the model maintains continuous transformation from input to output, significantly reducing processing time while preserving the logical flow through the network's layered architecture.

The patent merges independent parameter optimization into a unified end-to-end deep learning model where all parameters are jointly optimized through backpropagation. The model learns optimal transformations for head pose, gaze direction, and eyeball positioning simultaneously, ensuring that errors in one parameter do not accumulate from separate calculations but are instead coordinated through the shared neural network weights, thereby improving reliability.