Breaking down Polyblur: Fast Blind Correction of Small Anisotropic Blurs

Abstract

Polyblur is a two stage blind deblurring technique for removing small-sized blurs, like small camera shake or the lens point-spread function, proposed in 2021 by Delbracio et al. First, the blur is modeled with a zero-mean anisotropic Gaussian kernel whose parameters are rapidly estimated from the oriented blurry image gradients. Second, a sharp estimate is obtained by applying an approximate deconvolution filter, which is designed as a polynomial function of the estimated blurring kernel. Since in practice true blurs are not exactly Gaussian filters, the residual blur is gradually removed by repeating this two-stage procedure. Because it relies only on simple image manipulations, Polyblur is a quick blind deblurring technique, running in a fraction of a second on a smartphone. In this presentation, we analyze its key ingredients, showcase several use cases on real images, and provide Numpy and Pytorch implementations.

Implementing Handheld Burst Super-Resolution

“…In the companion demo shipped with this paper, we have to postprocess the fused image to display it side-by-side with the reference LR frame demosaicked with [19] and upsampled (if s is not set to 1) with bilinear interpolation. Starting from RGB images in the camera's color space, we use gamma compression with value of 2.2, the pixelwise S-curve x → 3x 2 − 2x 3 in guise of tone mapping, and lastly use the implementation of Polyblur from [7] to sharpen the image. The images are saved as 8-bit JPEG files with compression quality set to 95%.…”