Face recognition network construction method integrated with infrared image training

Abstract

The invention discloses a face recognition network construction method integrated with infrared image training. The method comprises the steps: packaging a convolution layer, a batch normalization layer, a correction linear unit layer and a depth separable convolution layer into a network single block; sequentially stacking the convolution layer, the batch normalization layer, the correction linear unit layer and the network single block from front to back to obtain network blocks, stacking four network blocks to obtain a backbone network, and sequentially inserting a random inactivation layerand a full connection layer behind the backbone network to obtain a network model; stacking and fusing the RGB image of the human face and the infrared image corresponding to the human face into a four-channel image, and inputting the four-channel image into a network model as training data; calculating a loss value between a real label of the training set sample and a prediction result output bythe full connection layer by utilizing a flexible maximum loss function of the additional angle interval; randomly initializing network parameters of the network model, and setting hyper-parameters;adopting a stochastic gradient descent method as a network optimization strategy, and repeatedly carrying out the calculation until the loss value converges to obtain an optimal network model.

Description

technical field [0001] The invention relates to the technical field of face recognition, in particular to a method for constructing a face recognition network fused with infrared image training. Background technique [0002] With the gradual advancement of the information age, various smart products have penetrated into people's surroundings, making people's daily life more and more convenient. However, along with this convenience, people also involve more exchanges of personal information. If personal information is lost, it will bring about economic loss and personal injury. Therefore, identity verification has become a process that can be seen everywhere now. Privacy Information. Among them, biometric technology is a commonly used method in the field of identity verification. It mainly identifies identities by analyzing and comparing human biometrics. Commonly used biometrics include faces, irises, fingerprints, etc., and facial features can be collected through non-cont...

AI Technical Summary

Problems solved by technology

However, these loss function formulas involve multiple hyperparameters, and a slight deviation will affect the performance of the model, requiring manual and careful debugging

[0007] It can be seen that most of the current deep learning-based algorithms use visible light images, and the performance of the algorithms will be affected by environmental factors, which will reduce the recognition accuracy, such as lighting changes, face occlusion, facial expressions, etc. There are some limitations when using

Images