eduzhai > Applied Sciences > Engineering >

Edge and Identity Preserving Network for Face Super-Resolution

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 33 pages

Abstract: Face super-resolution (SR) has become an indispensable function in securitysolutions such as video surveillance and identification system, but thedistortion in facial components is a great challenge in it. Moststate-of-the-art methods have utilized facial priors with deep neural networks.These methods require extra labels, longer training time, and largercomputation memory. In this paper, we propose a novel Edge and IdentityPreserving Network for Face SR Network, named as EIPNet, to minimize thedistortion by utilizing a lightweight edge block and identity information. Wepresent an edge block to extract perceptual edge information, and concatenateit to the original feature maps in multiple scales. This structureprogressively provides edge information in reconstruction to aggregate localand global structural information. Moreover, we define an identity lossfunction to preserve identification of SR images. The identity loss functioncompares feature distributions between SR images and their ground truth torecover identities in SR images. In addition, we provide aluminance-chrominance error (LCE) to separately infer brightness and colorinformation in SR images. The LCE method not only reduces the dependency ofcolor information by dividing brightness and color components but also enablesour network to reflect differences between SR images and their ground truth intwo color spaces of RGB and YUV. The proposed method facilitates the proposedSR network to elaborately restore facial components and generate high quality8x scaled SR images with a lightweight network structure. Furthermore, ournetwork is able to reconstruct an 128x128 SR image with 215 fps on a GTX 1080TiGPU. Extensive experiments demonstrate that our network qualitatively andquantitatively outperforms state-of-the-art methods on two challengingdatasets: CelebA and VGGFace2.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...