Motion Deblurring of Faces

This paper evaluates learning-based data-driven models for deblurring of facial images. Existing algorithms for deblurring, when used for facial images, often fail to preserve the facial shape and identity information. The best available models, which are used for general-purpose image deblurring, are pre-trained using only facial images. The Peak Signal to Noise Ratio (PSNR) Structural Similarity Index Measure (SSIM) and Time to deblur single images are the key metrics used for evaluating the models and for finding the most efficient model for deblurring facial images. From the results, the observation is that even though the PSNR value for DeblurGANv2 model is the highest, the best trade off between PSNR, SSIM, Time to deblur and visual quality is seen in DeblurGAN model.