Self-adaptive threshold selection method and face recognition method

Abstract

The invention discloses a self-adaptive threshold selection method and a face recognition method. The method comprises the steps: carrying out face quality evaluation of a face recognition test set in an actual scene, averagely dividing a face quality evaluation score into n intervals, dividing the face images in a query set into the corresponding intervals according to the face quality evaluation score, selecting proper face recognition similarity thresholds a and b for the face pictures of the n intervals according to actual requirements and the face recognition performance of the used algorithm, and establishing a corresponding table of the face quality scores and the face recognition similarity thresholds. The problem of face recognition similarity threshold setting problem for different-size and different-quality faces is solved by the refined self-adaptive face recognition similarity thresholds. The face recognition similarity threshold values of different face qualities can be set in a detailed manner, and then the overall face recognition effect is improved.

Description

technical field [0001] The invention relates to the fields of face recognition and deep learning, in particular to a method for selecting an adaptive threshold and a face recognition method. Background technique [0002] At present, face recognition algorithms based on deep learning are widely used in security fields such as smart check-in and smart prisons. The current face recognition technology takes the face image after face detection, face key point detection, and face correction as input, and obtains a feature vector through a deep convolutional neural network, and calculates the cosine of the feature vector and the bottom library feature vector Similarity, and then get the similarity between the face and the bottom library face. At this time, it is necessary to judge whether it matches according to the set threshold. In the process of realizing the present invention, the inventors have found that there are at least the following problems in the prior art: large-size ...

AI Technical Summary

Problems solved by technology

In the process of realizing the present invention, the inventors have found that there are at least the following problems in the prior art: large-size and high-quality face pictures have a large degree of discrimination between different types (that is, the degree of discrimination between different people is large and the degree of similarity is low), while The similarity of the same kind is high, and the face pictures of small-sized and low-quality faces have a small degree of heterogeneity discrimination, which shows that the use of a unified similarity threshold for face recognition does not take into account the difference in face quality

In actual scenarios, there is a big difference in the distribution of face quality. Using a unified similarity threshold for face recognition leads to large errors in accuracy and recall in unrestricted scenarios such as surveillance scenarios.

Images