Face recognition model training method and device, face recognition method and device, equipment and medium

Abstract

The invention relates to the technical field of model training, and provides a face recognition model training method and device, a face recognition method and device, equipment and a medium, and the method comprises: obtaining sample data which comprises a face image and a first identity vector corresponding to the face image; inputting the face image into a coding layer to obtain a first face feature vector; inputting a historical face feature vector and the first face feature vector in a preset vector queue into a vector compensation layer to obtain a second face feature vector; inputting the first face feature vector and the second face feature vector into a classification layer to obtain a second identity vector; determining whether the face recognition model converges according to the first identity vector, the second identity vector, the historical face feature vector and the historical identity vector; and if the face recognition model is not converged, storing the second face feature vector and the second identity vector into a preset vector queue to update the preset vector queue. The face recognition model trained by the scheme is more accurate.

Description

technical field [0001] The present application relates to the technical field of model training, and in particular, to a face recognition model training method, recognition method, apparatus, equipment and medium. Background technique [0002] Face recognition is a key technology for various biometric authentication applications such as electronic payment, smartphone screen lock, and video surveillance. The Convolutional Neural Networks (CNN) used in face recognition technology have greatly improved the recognition accuracy. However, many difficult problems of face recognition remain to be solved, for example, the mainstream of research introduces new objective functions to maximize the discriminability between classes and compactness within classes; they try to consider the identity representation vector by referring to All identities, this vector is the weight vector of the last fully connected layer used for identity classification. However, it is still difficult for tr...

AI Technical Summary

Problems solved by technology

However, it is still difficult for traditional methods to cover a large number of identities at one time, because these methods use mini-batch training (mini-batch), but the number of identities used is much less than the total number of identities due to memory constraints.

Examining tens of thousands of identities with a small mini-batch requires many iterations, which complicates the task of learning an optimal decision boundary in the embedding space while comprehensively considering all identities, which can be alleviated by increasing the mini-batch size. Certain issues, but usually due to memory constraints this solution is impractical and does not guarantee improved accuracy

Images