A training method for face representation models in realistic scenarios

Abstract

Face recognition has been widely used in daily life, but the existing model systems use processed high-quality datasets in training, while the face pictures in real scenes usually contain the influence of blurring, lighting, obscuring and other factors, thus making the existing face recognition models cannot perform well, and secondly, the existing face datasets have less data of Asian descent, resulting in the distribution learned by the models with the actual application. There is a certain error in the actual application. We propose a method to train face recognition models for realistic scenes by image augment of local face data to improve the classification accuracy of the models for low-quality images, and we demonstrate the feasibility of our method through experiments. Our method improves 0.619% and 0.414% in classifying images with added illumination and added random squares, respectively, compared to the current state-of-the-art methods.