Deep learning face verification method based on hybrid training

Abstract

Hybrid training based deep learning face verification method. Prepare the face data set; perform face and face key point detection on each image; normalize all faces to obtain a face image training set, and then divide it into training and verification data sets, and calculate all face images Mean image; mean training data set and verification data set obtained by subtracting the mean image from all face images; training deep convolutional neural network; generating corresponding triplets for each face image to form triplet training Data set and triplet verification data set; train the deep convolutional neural network again; detect the face and face feature points for the given two images to be verified, and subtract the mean image, and input it into the deep convolutional neural network , perform network feed-forward operation, and extract features; according to the selected threshold, when the distance between the extracted features of the two images is greater than the threshold, it is determined that the faces in the two images belong to the same person, otherwise they are determined to be different people.

Description

technical field [0001] The invention relates to face recognition in computer vision, in particular to a deep learning face verification method based on mixed training. Background technique [0002] Face recognition is a biometric identification method. Compared with other traditional biometric identification methods, face recognition has the advantages of non-contact, concealment, and high user acceptance. Face recognition is widely used in national security, security, access control and other fields, and has huge market value and scientific research value. Face recognition is an image-based recognition method. The challenge of image-based recognition methods is how to obtain effective feature representations from images for subsequent tasks such as recognition and classification. [0003] In the traditional face recognition method, the recognition task is decomposed into two independent parts of artificial feature design and classifier training, which are learned separat...

AI Technical Summary

Problems solved by technology

However, there are still some problems in the various methods based on the deep convolutional neural network model: First, the large demand for training data

It is sometimes difficult to obtain a large amount of training data.

Second, the time complexity of model training is high and the amount of calculation is large

For example, in FaceNet, in order to train a good model, it takes more than 2,000h of training time

Images