Deep joint super-resolution and feature mapping for low resolution face recognition

Abstract

To improve the accuracy in low resolution face recognition, a method based on super-resolution joint feature mapping is proposed. Firstly, a two-branch convolutional neural network is designed to extract features of high and low resolution face images. A super-resolution enhanced network cascading feature extraction network is used for feature mapping of low resolution face images. In this way, the high frequency information of low resolution image can be reconstructed, and features are extracted. Secondly, a fusion loss method is utilized, in which the loss of cosine and the image reconstruction are weighted and fusioned to increase the cosine similarity between image features of different resolutions. Finally, the experimental results based on FERET dataset validate that the test accuracy of two-branch framework is up to 98.2%, 99.1%, 99.5% with resolutions of 20× 20, 24× 24, and 36× 36 obtained by smooth downsampling. The proposed model outperforms up-to-date low resolution face recognition methods.