Method of masked face recognition by artificial intelligence technology

Patent metadata

Field	Value
Publication number	US-12562002-B2
Application number	US-202318447550-A
Country	US
Kind code	B2
Filing date	Aug 10, 2023
Priority date	Aug 12, 2022
Publication date	Feb 24, 2026
Grant date	Feb 24, 2026

Abstract

The present invention provides a method of masked face recognition from images by artificial intelligence technology including four steps: step 1: generating the images of face wearing a mask; step 2: training a deep learning model for face detection while wearing a mask; step 3: training the deep learning model for face feature extraction while wearing a mask; step 4: building a full pipeline of masked face recognition from images using the trained models from step 2, step 3, and some post-processing algorithms. The method aims to improve the accuracy of identity verification in the context of wearing masks has become popular and compulsory in life.

First claim

The invention claimed is:

1 . A method of masked face recognition from images by artificial intelligence technology comprising:
Step 1: generating the images of face wearing a mask; this step is performed based on a mask generation algorithm from normal face images, this step aims to generate training data for deep learning models; Step 2: training a deep learning model for face detection while wearing a mask; YOLO5Face model is proposed to use and is trained using the generated data from step 1, this step aims to increase the face detection accuracy while wearing a mask; Step 3: training the deep learning model for face feature extraction while wearing a mask; ArcFace loss function is proposed to use, and the model is trained using the generated data from step 1, this step aims to increase the face recognition accuracy while wearing a mask; Step 4: building a full pipeline of masked face recognition from images using the trained models from step 2, step 3, and some post-processing algorithms, further comprising: in step 4, a full pipeline of masked face recognition is built using the trained models from step 2, step 3 and some post-processing algorithms; first, the input images (from IP cameras, or devices such as tablets, personal computers, phones, cameras, etc.) are fed into the face detection model, which is the trained YOLO5Face model in step 2; the output returns the position of the faces in the images (both face wearing a mask and normal face without mask), and the detected faces are cut and aligned; after that, each of cut and aligned face image is fed into the face feature extraction model in step 3, the output returns the feature embedding vector of each face; next, the feature vector is normalized using the L 2 -norm algorithm; the following step is the process of searching the face in the database to determine the identity; this process uses the similarity searching algorithm from Faiss library (this is an open-source library); similarity searching is performed by calculating the similarity of the feature vector of a face image and all the feature vectors of the faces in the database; herein, the Euclidean distance algorithm (L 2 distance) is used to find the face in the database which has the maximum similarity to the input face (i.e., the distance is smallest); finally, a distance classification threshold is chosen to conclude that it is “the same person” or “different person”; if it is “the same person” then the last output returns the person's identity; in the opposite, if it is “different person” then the last output returns the result as “stranger”; based on many experiments, the proposed distance classification threshold is 1.01.

2 . The method of masked face recognition from images by artificial intelligence technology according claim 1 , further comprising:
in step 1, to enhance the data, the sample masks consist of medical mask, N95 mask, KN95 mask, and fabric mask, each type of mask consists of three images corresponding to three face positions: rotating left, rotating right and straight.

3 . The method of masked face recognition from images by artificial intelligence technology according claim 1 , further comprising:
in step 2, the deep learning model is proposed to use is YOLO5Face; this model can detect the faces from an image; train the model with the generated data from step 1; trained with the hyper-parameters including: the number of training epoch is 250, the learning rate is 10 −2 , the batch size is 256, the loss function is Cross-entropy, the optimization algorithm is Adam, and the weight decay is 0.0005 to avoid the over-fitting problem.

4 . The method of masked face recognition from images by artificial intelligence technology according claim 1 , further comprising:
in step 3, after the detected face is cut and aligned, another deep learning model is trained to learn the features of the face while wearing a mask, the output of the model is an embedded vector; the model is built based on the ResNet architecture as an embedded network and is trained using the Arcface loss function by the generated data from step 1; training with the hyper-parameters including: the number of training epoch is 30, the learning rate is 0.1, the batch size is 512, the loss function is Arcface, the optimization algorithm is Adam, and the weight decay is 0.0005 to avoid the over-fitting problem.

Assignees

• Viettel Group

Inventors

• Vu Thi Hanh
• Pham Van Muoi
• Nguyen Manh Quy

Classifications

• G06V40/161

  Detection; Localisation; Normalisation · CPC title

• G06V40/171

  Local features and components; Facial parts (eye characteristics G06V40/18); Occluding parts, e.g. glasses; Geometrical relationships · CPC title

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/776

  Validation; Performance evaluation · CPC title