Pose-variant face recognition using multiscale local descriptors
AssigneeMicrosoft Corporation
InventorsGang Hua, John Wright, Amir Akbarzadeh
Life of this patent
Estimating the term…
Abstract
Representing a face by jointly quantizing features and spatial location to perform implicit elastic matching between features. A plurality of the features are extracted from a face image and expanded with a corresponding spatial location in the face image. Each of the expanded features is quantized based on one or more randomized decision trees. A histogram of the quantized features is calculated to represent the face image. The histogram is compared to histograms of other face images to identify a match, or to calculate a distance metric representative of a difference between faces.
Claims20
A system for representing face images to account for pose variations during the comparison of the face images, said system comprising:
- a memory storing at least a face image;
- and
- a processor programmed to:
- access the face image stored in the memory;
- sample the accessed face image to generate vectors corresponding to features of the accessed face image;
- expand each of the generated vectors by a spatial location corresponding thereto in the accessed face image by concatenating pixel coordinates onto the features, the concatenation increasing a dimension of the features by at least two dimensions corresponding to the concatenated pixel coordinates;
- quantize each of the expanded vectors based on one or more randomized decision trees to enable implicit matching between one or more of the features, the implicit matching occurring due to quantizing each of the expanded vectors resulting from the concatenation, quantization of each of the expanded vectors capturing both deformations in feature value and image domain to generate a set of linear decision boundaries;
- calculate a histogram of the quantized expanded vectors to represent the accessed image;
- and
- associate the calculated histogram with the face image in the memory.
A method for representing an image of a face, said method comprising:
- accessing an image;
- extracting a plurality of features from the accessed image;
- expanding each of the plurality of extracted features with coordinates of a location corresponding thereto in the accessed image by concatenating the coordinates onto the extracted features, the concatenation increasing a dimension of the extracted features by at least two dimensions corresponding to the coordinates;
- quantizing each of the expanded features based on one or more randomized decision trees to enable implicit matching between one or more of the plurality of extracted features, the implicit matching occurring due to quantizing each of the features expanded by the concatenation of the coordinates onto the extracted features;
- and
- calculating a histogram of the quantized expanded features to represent the accessed image.
One or more computer-readable storage devices storing computer-executable components for comparing face images, said components comprising:
- a multiscale component that when executed by at least one processor causes the at least one processor to create a set of feature patches from an image of a face, said multiscale component further extracting a plurality of feature descriptors at a pre-defined pixel interval from the created set of feature patches;
- a match component that when executed by at least one processor causes the at least one processor to expand each of the features extracted by the multiscale component with coordinates of a location corresponding thereto in the image by concatenating the coordinates onto the features, said coordinates corresponding to the location from which each of the features are extracted, the concatenation increasing a dimension of the features by at least two dimensions corresponding to the coordinates, said match component further quantizing each of the expanded features based on one or more randomized decision trees to enable implicit matching, the implicit matching occurring due to jointly quantizing the features and the concatenated coordinates;
- and
- a similarity component that when executed by at least one processor causes the at least one processor to compare the quantized features corresponding to the image with quantized features corresponding to another image to generate a distance metric, said generated distance metric representing a difference between the image and the another image.
Dependencies are inferred from the claim text.