| Face super-resolution,which refers to the technique of reconstructing low-resolution face images into high-resolution face images,this technique provides reliable data support for various face-related applications by increasing the available information in face images.Recently,deep learning techniques,such as convolutional neural networks(CNN),have gained extensive attention in face super-resolution research,and have shown reconstruction performance that far exceeds that of traditional methods.However,as the research progresses,the face super-resolution models based on deep learning tend to be rather complicated and impractical to be deployed in real-world scenarios.In particular,when face feature information is introduced as supplementary information,such models put higher demands on computational resources,while undermining the real-time performance of subsequent face-related tasks.To address the above problems,this thesis investigates the lightweight design of face super-resolution networks and the lightweight introduction of face feature information based on pyramid reconstruction,and a lightweight reconstruction structure,respectively.The presented work includes:1.A lightweight face super-resolution network based on asymmetric U-pyramid reconstruction is proposed.The network has an asymmetric U-shaped structure,uses a coding and decoding structure to deepen the network,uses a coding structure composed of residual coding blocks for feature extraction,and introduces pyramid reconstruction in the decoding structure to achieve fast and accurate super-resolution.To reduce the number of parameters for up-sampling operations in the decoding block,a non-uniform channel widening strategy based on resolution selection is used.To avoid adding extra branches,face prior knowledge is introduced through heatmap loss.Experimental results show that this method lightly and effectively achieves super-resolution reconstruction of ultra-low resolution face images,reconstructs super-resolution face images with visual quality better than other current state-of-the-art methods at a lower model complexity,and has strong robustness.2.A lightweight face super-resolution network based on dual-pyramid reconstruction with face boundaries fusion is proposed.The network implements a lightweight feature extraction structure using Ghost Bottleneck and dense residual structure.A 5-category face boundary heatmap is introduced,and the face boundary heatmap and super-resolution face images are simultaneously predicted at each resolution level based on the dual pyramid reconstruction structure,and the predicted residual image similarity between the face image and the boundary heatmap is exploited to achieve a more lightweight prior extraction structure for face boundaries extraction using a more lightweight a priori extraction structure.The feature transformation-connection-fusion strategy is used to efficiently fuse the face feature maps and face boundary feature maps to achieve high-quality super-resolution reconstruction of faces.The experimental results show that the method we proposed is superior to other state-of-the-art method in terms of model complexity and various evaluation metrics. |
