Deep HDR Reconstruction of Dynamic Scenes

High dynamic range reconstruction of dynamic scenes from several images of different exposures is a challenging problem. In state-of-the-art methods, several researchers try to fix this problem by proposing traditional algorithm such as patch-based methods and motion rejection methods, while others attempt to formulate HDR reconstruction as a deep learning model. In our paper, we put forward an approach combining both traditional pipeline and deep learning algorithm, in which we align the input LDR images to the reference one using an optical flow method [1] based on deep convolution network and then merge aligned LDR images into HDR images using a powerful network which we call MergeNet. MergeNet is able to not only remove alignment artifacts but also produce obvious HDR effect. We perform extensive experiments to demonstrate that our pipeline and deep network produce high-quality HDR images and outperform the state-of-the-art approaches on challenging dynamic scenes.