InvHDR: Inverse Tone Mapping With Invertible Neural Network

In general, many cameras capture a relatively small range of luminosity due to the price of high-performance camera sensors. We can compensate this by creating high dynamic range (HDR) images using multi-stack fusion, which combines information from multiple low dynamic range (LDR) images of various dynamic ranges. However, ghost artifact is inevitably caused by moving objects when acquiring multiple images. Therefore, many state-of-the-art methods create pseudo-image stack of multiple dynamic ranges from one single LDR image and fuse them to a HDR image. In this paper, we use invertible neural network to improve the existing methods and propose a conditional network structure with improved performance on multiple indicators.