<p>With the development of convolutional neural network (CNN), face super-resolution achieves great improvement in terms of both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). However, most of the existing methods still have poor performance in high-level visual tasks such as face recognition. Besides, most CNN-based methods usually rely on only one reconstructed image, so it is difficult to ensure stable and accurate high-frequency details. To solve the issues, we propose a tree-based multi-branch face super-resolution network driven by prior knowledge and recognition task (Face-SRMN). Tree-based structure is designed to build a multi-branch super-resolution network, which can enhance feature representation capability and enrich high-frequency details of the restored image. Dual-channel Residual structure is proposed for basic blocks, which lets more low frequencies pass. The basic block incorporating dense residual structure and attention mechanism is further designed. Combined with face prior information module, it can enhance the feature extraction of face information and enable the network to perform global information adaptive adjustment in both channels and space while deepening the network. Furthermore, a cascaded face recognition network is introduced to jointly train the network towards improving face recognition accuracy by combining the losses of super-resolution network and face recognition network. Experimental results show that the proposed method performs favorably against the state-of-the-art super-resolution algorithms in term of visual quality, PSNR, SSIM and face recognition accuracy.</p>