US20220400246A1 테스트

The prior art discloses gaze correction by identifying eye locations, extracting SIFT descriptors as local image features, and applying morphological transformations to replacement eye imagery. However, it does not disclose multi-view images, deriving a single feature vector from plural local descriptors, or using that feature vector to look up reference displacement vector fields. The prior art uses a database of replacement imagery rather than displacement vector fields looked up by feature vectors.

The prior art discloses gaze correction of eye regions in images, including detecting eye regions, computing displacements for eye outer points, and warping the eye texture. However, it does not disclose multi-view images, feature vectors derived from plural local image descriptors, or lookup of reference displacement vector fields using such feature vectors.

US9538130B1 discloses gaze correction by identifying eye regions in video frames and substituting them with stored forward-looking eye portions. However, it does not disclose multi-view images, feature vectors derived from local image descriptors, displacement vector fields, or lookup of reference displacement vector fields. The approach is fundamentally different: substitution-based rather than transformation/displacement-based.

US9384384B1 discloses adjusting facial representations in images using a processor and involves face/feature identification, but does not specifically address multi-view images, gaze correction, eye patch identification, feature vectors from local image descriptors, displacement vector fields, or lookup of reference displacement vector fields. The prior art operates on a fundamentally different approach using manifolds for face model generation rather than the claimed feature-vector-to-displacement-vector-field lookup approach for gaze correction.

CN107534755B discloses gaze correction in video frames of a user's face, which broadly overlaps with the preamble concept of adjusting images to correct gaze. It also involves identifying/tracking the user's face and eyes. However, it does not disclose multi-view images, feature vectors derived from plural local image descriptors, lookup of reference displacement vector fields, or transforming image patches using displacement vector fields. The core technical approach of the target patent (descriptor-based feature vectors, displacement vector field lookup and application) is entirely absent from this prior art.

US8467627B2 discloses detecting eyes in a face area and performing warp processing on an image based on detected eye positions. However, it does not disclose multi-view images, gaze correction, feature vectors from local image descriptors, displacement vector fields derived by looking up reference data, or transforming eye patches in accordance with displacement vector fields. The prior art's warp processing is a simple non-linear warp (e.g., fisheye lens filter) centered on detected eyes, not a gaze-correction technique using learned displacement vector fields.

US9633250B2 is directed to facial landmark localization using globally aligned regression. It deals with face images, feature extraction at landmark locations, and regression to update landmark positions. However, it does not address multi-view images, gaze correction, eye-specific image patches, displacement vector fields for image transformation, or lookup of reference data comprising reference displacement vector fields. The overlap is limited to general concepts of processing face images with a processor and extracting features.

JP5645450B2 relates to multi-view image processing using multiple cameras from different viewpoints, but it is focused on viewpoint interpolation and blur processing for smooth virtual camera movement. It does not address gaze correction, eye detection, feature vectors from local image descriptors, displacement vector fields, or any lookup-based transformation for correcting eye regions. The only overlap is the general concept of processing multi-view images.