A Walk Through Image Deblurring: From Model-Based to Generative Restoration

Image deblurring has seen tremendous progress in recent years mostly coming hand-in-hand with the success of deep neural networks. Greater computational power, reliable and accessible training frameworks and large amounts of data have enabled deep image processing models that exceed or are on par with those conceived through careful and artisan modeling. During this talk I will present our recent work on image deblurring with a focus on two distinct scenarios. First, I will introduce Polyblur, a highly efficient blind restoration method for removing mild blur in natural images. Polyblur estimates slight image blur and compensates for it by combining multiple applications of the estimated blur allowing processing of a 12MP photo on a modern mobile phone in a fraction of a second. In the second part of the talk, I will discuss how to train deep image enhancement models for improved realism in restored images. I will present an alternative approach using a conditional diffusion model to stochastically refine the output of a deterministic predictor capable of producing realistic results. To conclude this talk, I will showcase the newly introduced Unblur feature in the Google Pixel 7 Pro.

In this talk, we will discuss how human perception should be leveraged to further address problems within speech enhancement, and we will discuss how human perception can be predicted and used to improve noise reduction. Additionally, we will provide highlights of my work in complex-domain speech enhancement, which encouraged full signal reconstruction.