Face recognition is a modern security solution that is quickly and easily integrated into most existing devices, so this system is widely applied to several domains as one of the security authorizations. Developing face recognition models using mainstream architectures (AlexNet, VGGNet, GoogleNet, ResNet, and SENet) will make it difficult to implement the models on mobile devices and embedded systems. In addition, low resolution images, such as those from CCTV surveillance cameras or drones, pose challenges for the models to recognize faces, as the images lack sufficient details for identification. Therefore, this research aims to analyze the performance of a face recognition model developed using the lightweight VarGFaceNet architecture with the adaptive margin loss AdaFace on a low-resolution image dataset. Based on the evaluation results on the LFW dataset, an accuracy of 99.08% was achieved on high-resolution data (112x112 pixels), while on the lowest synthetic low-resolution data (14x14 pixels), an accuracy of 79.87% was obtained with the assistance of the Real-ESRGAN and GFP-GAN super-resolution models. On the TinyFace dataset, without fine-tuning, a Rank-1 accuracy of 46.08% was achieved without using super-resolution models and 45.03% when utilizing super-resolution models.
Copyrights © 2023