Method and apparatus for determining image quality

Patent metadata

Field	Value
Publication number	US-11487995-B2
Application number	US-201816050346-A
Country	US
Kind code	B2
Filing date	Jul 31, 2018
Priority date	Sep 29, 2017
Publication date	Nov 1, 2022
Grant date	Nov 1, 2022

Abstract

Embodiments of the present disclosure disclose a method and apparatus for determining image quality. The method comprises: acquiring a to-be-recognized image and facial region information used for indicating a facial region in the to-be-recognized image; extracting a face image from the to-be-recognized image on the basis of the facial region information; inputting the face image into a pre-trained convolutional neural network to obtain probabilities of each pixel comprised in the face image belonging to a category indicated by each category identifier in a preset category identifier set; inputting the face image into a pre-trained key face point positioning model to obtain coordinates of each key face point comprised in the face image; determining a probability of the face image being obscured on the basis of the probabilities and the coordinates; and determining whether the quality of the face image is up to standard on the basis of the probability.

First claim

What is claimed is:

1. A method for determining image quality, comprising:
acquiring a to-be-recognized image and facial region information obtained by performing facial recognition on the to-be-recognized image in advance and used for indicating a facial region in the to-be-recognized image;
extracting a face image from the to-be-recognized image on the basis of the facial region information;
inputting the face image into a pre-trained convolutional neural network to obtain probabilities of each of pixels comprised in the face image belonging to a category indicated by each category identifier in a preset category identifier set, the convolutional neural network being used to represent a corresponding relationship between an image containing a face and a probability of a pixel belonging to a category indicated by a category identifier in the category identifier set, and the category identifier set having a category identifier indicating a category representing a face part, wherein the convolutional neural network is trained by:
collecting a preset training sample comprising a sample image displaying a sample face and annotations of the sample image, wherein the sample face comprises a plurality of sample face parts, and each of the annotations comprising a data flag for representing whether a pixel in the sample image belongs to the category indicated by the category identifier in the category identifer set; and
training using a machine learning method in the basis of the sample image, the annotation, a preset classification loss function and a back propagation algorithm to obtain the convolutional neutral network, the classification loss function being used for representing a degree of difference between a probability output by the convolutional neural network and the data flag comprised in the annotation;

inputting the face image into a pre-trained key face point positioning model to obtain coordinates of each key face point comprised in the face image, the key face point positioning model being used for representing a corresponding relationship between an image containing a face and coordinates of each key face point;
determining a probability of the face image being obscured on the basis of the probabilities and the coordinates; and
determining whether the quality of the face image is up to standard on the basis of the probability of the face image being obscured.

2. The method according to claim 1 , wherein the convolutional neural network comprises five convolutional layers and five deconvolutional layers, the convolutional layers being used for downsampling inputted information with a preset window sliding step, and the deconvolutional layers being used for upsampling the inputted information with a preset amplification factor.

3. The method according to claim 2 , wherein the window sliding step is 2, and the amplification factor is 2.

4. The method according to claim 1 , wherein the determining a probability of the face image being obscured on the basis of the probabilities and the coordinates comprises:
inputting the probabilities of the each pixel comprised in the face image belonging to the category indicated by the each category identifier in the category identifier set and the coordinates of the each key face point comprised in the face image into a preset probability calculation model to obtain the probability of the face image being obscured, wherein the probability calculation model is used to represent a corresponding relationship between inputted information and a probability of a face image being obscured, and the inputted information comprises: probabilities of each pixel comprised in an image containing a face belonging to a category indicated by each category identifier in the category identifier set and coordinates of each key face point comprised in the image.

5. The method according to claim 1 , wherein the determining a probability of the face image being obscured on the basis of the probabilities and the coordinates further comprises:
determining a face part region set on the basis of the coordinates;
determining, for each pixel comprised in the face image, a category indicated by a category identifier corresponding to a maximum probability corresponding to the pixel as a category the pixel attributed to;
calculating, for each face part region, a probability of the face part region being obscured on the basis of a category each pixel comprised in the face part region attributed to; and
determining the probability of the face image being obscured on the basis of probabilities of each face part region in the face part region set being obscured.

6. The method according to claim 5 , wherein the calculating, for each face part region, a probability of the face part region being obscured on the basis of a category each pixel comprised in the face part region attributed to comprises:
determining, for the each face part region, a number of pixels in the face part region attributed to a category not matching a face part represented by the face part region, and determining a ratio of the number to a total number of pixels comprised in the face part region as the probability of the face part region being obscured.

7. The method according to claim 5 , wherein the calculating, for each face part region, a probability of the face part region being obscured on the basis of a category each pixel comprised in the face part region attributed to further comprises:
determining, for the each face part region, a target pixel group comprising a target pixel, in the face part region, attributed to a category not matching a face part represented by the face part region, summing probabilities of each target pixel in the determined target pixel group belonging to a category matching the face part to obtain a first value, summing probabilities of each pixel in the face part region belonging to a category matching the face part to obtain a second value, and determining a ratio of the first value to the second value as the probability of the facial region being obscured.

8. The method according to claim 5 , wherein the determining the probability of the face image being obscured on the basis of probabilities of each face part region in the face part region set being obscured comprises:
determining the probability of the face image being obscured on the basis of an average of the probabilities of the each face part region in the face part region set being obscured.

9. The method according to claim 5 , wherein the determining the probability of the face image being obscured on the basis of probabilities of each face part region in the face part region set being obscured further comprises:
acquiring a preset weight for representing an importance level of a face part; and
weighting and summing the probabilities of the each face part region in the face part region set being obscured on the basis of the weight to obtain a numeric value, and defining the numeric value as the probability of the face image being obscured.

10. The method according to claim 1 , wherein the extracting a face image from the to-be-recognized image on the basis of the facial region information comprises:
enlarging a range of the facial region indicated by the facial region information to obtain a first facial region; and
extracting the first facial region to obtain the face image.

11. The method according to claim 10 , wherein the facial region is a rectangular region; and
the enlarging the range of the facial region indicated by the facial region information to obtain a first facial region comprises:
increasing a height and width of the facial region indicated by the facial region information by a preset multi

Assignees

• Baidu online network technology beijing co ltd

Inventors

• Du Kang

Classifications

• G06V10/993

  Evaluation of the quality of the acquired pattern · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06N3/048

  Activation functions · CPC title