Scalable training data preparation pipeline and efficient distributed trainer for deep neural networks in facial recognition
AssigneeClearview AI, Inc.
InventorsZhi Liu, Cam-Hoan Ton-That
Life of this patent
Estimating the term…
Abstract
This disclosure provides a highly scalable training data preparation pipeline for data cleaning and augmentation with the aim of extracting the most meaningful information while keeping the noise level low, as well as a highly efficient distributed trainer for the deep neural networks suitable for facial recognition. The goal is to train deeper and larger neural networks with larger and higher quality facial image datasets iteratively and frequently without incurring prohibitive costs and drastic delays.
Claims24
A method for generating a facial image dataset for training a facial recognition model, comprising:
- receiving a set of facial images comprising a plurality of subsets of facial images, each of the subsets of facial images is associated with a pre-assigned identity label;
- performing intra-identity facial image deduping for each subset of facial images by:
- determining pairwise intra-identity facial image similarity scores between the facial images in the each subset of facial images;
- identifying two or more facial images within the each subset of facial images as duplicate facial images associated with the same identity label if the two or more facial images have an intra-identity similarity score greater than or equal to an intra-identity similarity threshold value;
- and
- retaining only one of the duplicate facial images after removing the rest of the duplicate facial images;
- performing intra-identity clustering by applying a first clustering similarity threshold to group the deduped facial images in the each subset of facial images into a plurality of clusters;
- and
- performing intra-identity second-order clustering by determining pairwise cluster similarity scores between centroid vectors of clusters and combining clusters having pairwise cluster similarity score greater than or equal to a second clustering similarity threshold to obtain an intra-identity cleaned facial image dataset, wherein the combined clusters and remaining uncombined clusters are respectively associated with a plurality of actual identities.
A system for generating a facial image dataset for training a facial recognition model, comprising:
- a non-transitory, computer-readable memory;
- one or more processors;
- and
- a computer-readable medium containing programming instructions that, when executed by the one or more processors, configure the system to:
- receive a set of facial images comprising a plurality of subsets of facial images, each of the subsets of facial images is associated with a pre-assigned identity label;
- perform intra-identity facial image deduping for each subset of facial images by:
- determining pairwise intra-identity facial image similarity scores between the facial images in the each subset of facial images;
- identifying two or more facial images within the each subset of facial images as duplicate facial images associated with the same identity label if the two or more facial images have an intra-identity similarity score greater than or equal to an intra-identity similarity threshold value;
- and
- retaining only one of the duplicate facial images after removing the rest of the duplicate facial images;
- perform intra-identity clustering by applying a first clustering similarity threshold to group the deduped facial images in the each subset of facial images into a plurality of clusters;
- and
- perform intra-identity second-order clustering by determining pairwise cluster similarity scores between centroid vectors of clusters and combining clusters having pairwise cluster similarity score greater than or equal to a second clustering similarity threshold to obtain an intra-identity cleaned facial image dataset, wherein the combined clusters and remaining uncombined clusters are respectively associated with a plurality of actual identities.
Dependencies are inferred from the claim text.