Defocus Map Estimation and Blur Removal from a Single Dual-Pixel Image

Ioannis Gkioulekas
Jiawen Chen
Neal Wadhwa
Pratul Srinivasan
Rahul Garg
Shumian Xin
Tianfan Xue
International Conference on Computer Vision (2021)
Google Scholar

Abstract

We present a method to simultaneously estimate an image's defocus map, i.e., the amount of defocus blur at each pixel, and remove the blur to recover a sharp all-in-focus image using only a single camera capture. Our method leverages data from dual-pixel sensors that are common on many consumer cameras. Though originally designed to assist camera autofocus, dual-pixel sensors have been used to separately recover both defocus maps and all-in-focus images. Past approaches have solved these two problems in isolation and often require large labeled datasets for supervised training. In contrast with those prior works, we show that the two problems are connected, model the optics of dual-pixel images, and set up an optimization problem to jointly solve for both. We use data captured with a consumer smartphone camera to demonstrate that after a one time calibration step, our approach improves upon past approaches for both defocus map estimation and blur removal, without any supervised training.

Research Areas