Low Resolution Face Recognition With Deep Learning

With the rapid development of computer vision,face recognition has evolved from simple backgrounds to complex scenes,such as various poses,lighting,expressions,noise,occlusion,makeup,age,race,gender,and so on.Various existing face recognition systems assume a high recognition rate under some constrained environments.In the realistic environment,especially in video surveillance applications,the images captured by the camera are blurred and have low resolution,leading to a reduction in recognition accuracy.Therefore,how to improve the accuracy of face recognition under low-resolution conditions is an important research direction.This thesis addresses three key issues on the low-resolution face recognition,including low-resolution face image processing methods,the applicability of ordinary face recognition methods to low-resolution face recognition and methods for improving low-resolution face recognition accuracy.A series of modifications are made to improve the performance of low-resolution face recognition systems.The main contributions can be summarized as follows.(1)The state-of-the-art face recognition methods at home and abroad are briefly reviewed,and the main problems and difficulties of low-resolution face recognition are provided.(2)A new low-resolution face recognition method based on modular constraint CentreFace is proposed.The CentreFace method can reduce the distance within the feature class due to the addition of the center loss function,but at the same time,the distance between classes is also decreasing.Therefore,the benefit of center loss function to the performance of the low-resolution face recognition model is not obvious.To remedy this problem,a modular loss function is proposed.The modular loss function can increase the distance between classes while maintaining the distance within the feature class and improving the generalization ability of the model.After conducting 4sets of analysis and comparison experiments on the QMUL-Surv Face data set,we show that the proposed method performs better than the CentreFace algorithm.(3)A low-resolution face recognition method based on knowledge distillation is proposed.This method is based on the combination of Li-Arc Face and knowledge distillation.Knowledge distillation enables the low-dimensional features from student networks to inherit knowledge from teacher models and training data.At the same time,this method eliminates some interference in high-dimensional features and pays more attention to discovering the similarity of human faces.Experiments show that the improved performance can be achieved.