A Face Recognition Method Based on Transfer Learning and Attention Mechanism

Abstract

With the popularization of network technology and the development of informatization, the security of personal identity information is receiving increasing attention. In fields such as finance, healthcare, and security, security requirements are increasing, and more reliable and secure identity verification methods are needed to protect personal information from illegal acquisition and use. Traditional identity verification methods, such as passwords and PIN codes, face problems such as being guessed, forgotten, and stolen. Based on the understanding of the limitations and shortcomings of traditional identity verification methods, biometric recognition technology has emerged. At present, it mainly includes biometric recognition technologies such as fingerprint recognition[1], voice recognition[2], facial recognition[3], iris recognition, retinal recognition[4], and DNA recognition[5] . As an important branch, facial recognition technology has shown tremendous potential and advantages in the field of information security In order to study the performance of facial recognition in small sample scenarios, the ECANet attention mechanism was introduced into the classic ResNet50 network, and a new network model, ResNet50-ECA[6], was constructed. Firstly, a small-scale face SLFW dataset containing only 28 classifications was created on the famous LFW (Labeled Faces in the Wild) dataset[7]. When processing these facial images, an advanced image enhancement technique, namely adaptive histogram equalization with limited contrast, was adopted. Effectively enhances the contrast between light and dark in the image, making facial features clearer and more prominent. In order to expand the dataset and enhance the robustness of the model, data augmentation was also performed. This includes random rotation and horizontal inversion of images, which can generate more diverse and rich training samples. Pre trained network models on the ImageNet dataset with parameters and weights, then fine tuned these models and applied them to a small sample face dataset.