Background: HDR PhotographyThe goal of HDR imaging is to have a greater dynamic range between the darkest and lightest areas of an image.HDR images can represent more precisely real scenes intensity levels.Mainly, there are three common techniques for HDR photography: (i) sequentially capturing and fusing multiple exposures images (e.g. , ). While this method is easily supported by existing cameras, additional de-ghosting and motion stabilization techniques are usually needed. (ii) Simultaneously utilizing multiple sensors to capture different exposures (e.g. ), this sophisticated approach is more expensive and need rigorous calibration. (iii) Capturing single codded exposure image, along with a proper algorithm for HDR Image reconstruction (, , and ).
ContributionIn this project, we propose using deep learning to reconstruct HDR image from single shot coded pixel exposure computational cameras. We used deep networks for coded mask calibration, and HDR image reconstruction. Our method supports joint designing of optical elements and learning algorithms. Once trained, the obtained deep networks weights can be used for quick and simple coded mask calibration and a well HDR reconstruction of images greatly differing from the training set.