Face Super-Resolution Network Based on Efficient Attention Mechanisms
Currently,deep learning methods based on facial priors have been able to better recover the details of degraded face images.However,these methods still have some shortcomings.On the one hand,multi-task joint training requires additional prior labeling on the dataset,and the introduction of the priors will significantly increase the computational cots of the network.On the other hand,most methods only use convolutional neural networks(CNN),the limited perceptive field of CNN will reduce the in-tegrity and accuracy of the reconstructed facial images.To address these problems,a face super-resolution network based on effi-cient attention mechanisms is proposed.The network combines CNN with several attention mechanisms such as Transformer.It can effectively recover the global structure and local texture details of a face without the aid of any facial priors.Extensive experiments on various datasets show the effectiveness of the proposed network.
face super-resolutionTransformerconvolutional neural networkattention mechanism
国家科技期刊平台
NSTL
万方数据
