Face recognition algorithm training method based on multi-objective learning

Abstract

The invention discloses a face recognition algorithm training method based on multi-objective learning, comprising the following steps: randomly initializing neural network parameters, using a loss function based on face identity and a loss function based on face feature point position constraint to minimize the learning objective to train the deep convolution network; when the prediction accuracyrate of the face identity reaches the threshold, calculating the loss function based on the distance within the face feature classes and the loss function based on the distance between the face feature classes, and performing calculation of the loss function based on the face identity and the loss function of the face feature point position constraint on each sample; based on the artificial weight setting, weighting each loss function to obtain the total loss function, and realizing the back propagation based on the total loss function. Therefore, the update of the network parameters is realized, and the network training is stopped when the accuracy is stable to get the trained face recognition model.

Description

technical field [0001] The invention relates to the technical field of face recognition algorithm training, in particular to a face recognition algorithm training method based on multi-objective learning. Background technique [0002] The face recognition algorithm based on deep convolutional network is the mainstream model of the current face recognition algorithm. The main technologies include the use of multi-layer convolutional neural network, the activation function based on ReLU (rectified linear unit, Rectified linear unit) and the activation function based on TripLet The supervised training method of loss has two main shortcomings: [0003] 1. The image is only initially aligned when the training image is input, and the influence of the face pose is not considered during the training process, and the robustness to face pose changes is limited; [0004] 2. The training goal is to target the prediction accuracy rate of the samples in the training set. In actual deploy...

AI Technical Summary

Problems solved by technology

[0003] 1. The image is only initially aligned when the training image is input, and the influence of the face pose is not considered during the training process, and the robustness to face pose changes is limited;

[0004] 2. The training goal is to target the prediction accuracy rate of the samples in the training set. In actual deployment and application, the features extracted based on the deep convolutional network require that the features corresponding to different face images of the same identity are close to each other, while the features of people with different identities are close to each other. The features corresponding to the face image are far away from each other, and there is a deviation in the target of the algorithm training and application, which leads to the limitation of the generalization ability of the recognition algorithm

Images