A facial motion deblurring algorithm based on semantics and GAN

Face deblurring is a subtask of image deblurring. Non-deep learning deblurring methods estimate the blur kernel based on mathematical statistics, and then get a sharp image from the degraded image through deconvolution, which is not suitable for most real-life situations. General deep learning deblurring methods often have poor visual effects when applied to face deblurring, and many details on the face cannot be reproduced. Therefore, face deblurring requires some prior information based on the face structure. The progressive semantic face deblurring generative adversarial network (PSFD-GAN) proposed in this article uses the face semantic map provided by the CelebAMaskHQ dataset as a priori information to supervise network training, and adopted a progressive learning strategy to gradually deblur from coarse to fine. The final experimental results on the two data sets CelebA and FFHQ showed the effectiveness of network deblurring, and the visual effects and indicators were better than previous work.