Neural-Augmented HDR Imaging via Two Aligned Large-Exposure-Ratio Images

Two aligned large-exposure-ratio (LER) images are captured simultaneously by emerging devices for a high dynamic range (HDR) scene to avoid ghosting artifacts from appearing in an image fused from them. However, brightness order reversal (BOR) usually appears when the two LER images are directly fused together by an existing multiscale exposure fusion (MEF) algorithm. In this article, a novel neural augmentation-based exposure interpolation framework is introduced in this article by seamlessly integrating physics-driven and data-driven approaches. The physics-driven method infers an initial representation while the data-driven one learns the remaining information for the middle-exposure image resulting from the two LER images. They compensate each other well in the proposed framework. These two LER images and the interpolated image are fused together using a novel data-driven MEF algorithm which can further interpolate information for the two LER images. Experimental results show that the proposed framework can indeed solve the BOR problem for the fusion of two LER images. Real high-frequency information from the two LER images is also preserved well in the fused image.