Convolutional neural network training method, face recognition method and device

Abstract

The invention discloses a convolutional neural network training method, a face recognition method and a device. By adopting a first joint training supervision function composed of a cross-entropy loss function and a normalized comparison loss function with two thresholds, According to the eigenvectors of the face images in each sample in the source domain training sample set, the first error is determined, and the network parameters of the convolutional neural network are adjusted through the first error. Among them, the first threshold is used to compare with the Euclidean distance of the feature vectors of the two face images in the positive sample pair, and the second threshold is used to compare with the Euclidean distance of the feature vectors of the two face images in the negative sample pair, In this way, both the supervised training of negative sample pairs and the supervised training of positive sample pairs can be controlled, and the training efficiency and accuracy of CNN can be improved. Therefore, when the trained CNN is applied to the face recognition method, the generalization ability of the face recognition method can be improved.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a training method of a convolutional neural network, a face recognition method and a device. Background technique [0002] At present, in order to better recognize face images, more and more recognition processes need to use Convolutional Neural Network (CNN). In order to ensure accurate recognition results, CNN needs to be trained repeatedly. At present, the contrastive loss function (Contrastive Loss) with a single threshold is generally used to train CNN. However, since this contrastive loss function only controls the convergence of the negative sample pairs in the sample set by a single threshold to train the CNN, it does not control the convergence of the positive sample pairs in the sample set. Since there are a large number of positive sample pairs that are easy to converge in the sample set, and these positive sample pairs are not supervised for training, ...

AI Technical Summary

Problems solved by technology

However, since this contrastive loss function only controls the convergence of the negative sample pairs in the sample set by a single threshold to train the CNN, it does not control the convergence of the positive sample pairs in the sample set

Since there are a large number of positive sample pairs that are easy to converge in the sample set, and these positive sample pairs are not supervised for training, it not only increases the burden of CNN, but also hinders CNN from supervised learning of other positive sample pairs that are not easy to converge, making CNN The training efficiency and accuracy are reduced, which leads to less than ideal training results of CNN

Images