A face image quality classification and evaluation method based on supervised deep learning

Abstract

The invention discloses a face image quality classification and evaluation method based on supervised deep learning. The method comprises the steps of screening out a high-definition image set Q2 froman original image set Q0; carrying out face recognition tests under different brightnesses, and selecting a corresponding brightness value when the recognition rate is obviously changed to obtain a high-definition image set Q3 with proper brightness; carrying out face recognition testing on a plurality of images of the same object in Q3, and selecting an image which exceeds a certain threshold value and has a high similarity score as a standard image set A; the Q1 and the A are subjected to face recognition testing, and the image is divided into a very fuzzy image which may generate a poor authentication result, a relatively clear image which may generate a good authentication result in most environments and a very clear image which may generate a good authentication result in any environment according to the similarity score; Each type of image is subdivided into an image with appropriate brightness, an image with slightly bright brightness and an image with slightly dark brightness;and finally, the face images are divided into nine types.

Description

technical field [0001] The invention relates to the technical field of face recognition, in particular to a method for classifying and evaluating the quality of face images based on supervised deep learning. Background technique [0002] Face image quality is a key factor related to the accuracy of face recognition. Refusing to authenticate or recognize low-quality face images will significantly improve the accuracy rate, so the evaluation of face image quality is particularly important. To evaluate the quality of face images with supervised deep learning methods, it is necessary to establish target quality labels for face images. [0003] At present, there are three main ways to establish target quality labels for face images: [0004] The first is to combine various measurements of image quality factors such as sharpness, brightness, etc. into a single value that reflects the overall quality of the face. So far, this approach has not had much success, and the results of ...

AI Technical Summary

Problems solved by technology

Score-based quality is hard to predict due to specific matchers and subtle differences in pairwise quality factors (i.e. comparison score is a function of two face images, but scores are used to mark the quality of individual face images)

The comparison score is obtained from a pair of images, so labeling individual images based on the comparison score (or performance) is problematic. On datasets with different distortions, the correlation between the observer's subjective rating and the algorithm's rating is very low. Aspects are not very consistent

Images