Efficient Deep Attentive Pixels Network in Face Super-Resolution at Scale Factor of 16

Abstract

Nowadays, Face Super-Resolution (FSR) models utilize the fusion approach, which combines the attention technique with the super-resolution network. The fusion approach has been proposed and solves the problem of FSR. Facial attributes have been effectively used to guide low-level information of the face to perform robust face image restoration. Iterative techniques appraised the value of facial landmarks to boost the reconstruction capability of the super-resolution network. Nevertheless, the network parameters in FSR are high, while the learning rate is still low. This paper proposes an attention mechanism combined with the Face Alignment Network (FAN). The proposed attention mechanism consists of channel attention and a non-local module. Our proposed model outperforms at the scale of $\times 16$ compared to the other state-of-the-art models.