Low-Resolution Face Recognition Method Combining Super-Resolution and Improved DCR Model

Low-resolution face recognition is a challenging problem in the face recognition algorithm. Most face recognition algorithms have a significant reduction in recognition accuracy when facing low-resolution faces. In this paper, we propose a DCR(Deep Coupled ResNet) model based on super-resolution reconstruction, named SR-DCR model, to tackle this challenging. Firstly, we designed a super-resolution face reconstruction model based on the facial feature. The "perceptual loss + classification loss" is used as the loss function of the reconstructed network. And the facial feature is used to constrain the reconstruction process to make the reconstructed face more favorable for recognition. Then a series of improvements and adjustments were made to the DCR model. Improved the network structure of the main network, and its cross-layer connection method was changed to a close connection, which further promoted feature propagation in the network and enhanced feature reusability; The loss function of the main network is improved, and the improved "triplet loss" is used as a loss function to directly optimize the facial feature between high resolution and low-resolution, further enhancing the high and low resolution. The aggregation between facial features; at the same time, the training method of the DCR model is adjusted, and the steps of joint training between the main network and branch network are added to further optimize DCR network parameters. Finally, the super-resolution reconstruction and the improved DCR model are combined to recognize low-resolution faces. The performance of the proposed algorithm is compared with that of other advanced methods on LFW and SCFace datasets. Experimental results show that the algorithm improves the recognition rate by 3.5% based on DCR. Effectively improve the accuracy of face recognition at low resolution.