Single Image HDR Reconstruction Using a CNN with Masked Features and Perceptual Loss

Digital cameras can only capture a limited range of real-world scenes'luminance, producing images with saturated pixels. Existing single image highdynamic range (HDR) reconstruction methods attempt to expand the range ofluminance, but are not able to hallucinate plausible textures, producingresults with artifacts in the saturated areas. In this paper, we present anovel learning-based approach to reconstruct an HDR image by recovering thesaturated pixels of an input LDR image in a visually pleasing way. Previousdeep learning-based methods apply the same convolutional filters onwell-exposed and saturated pixels, creating ambiguity during training andleading to checkerboard and halo artifacts. To overcome this problem, wepropose a feature masking mechanism that reduces the contribution of thefeatures from the saturated areas. Moreover, we adapt the VGG-based perceptualloss function to our application to be able to synthesize visually pleasingtextures. Since the number of HDR images for training is limited, we propose totrain our system in two stages. Specifically, we first train our system on alarge number of images for image inpainting task and then fine-tune it on HDRreconstruction. Since most of the HDR examples contain smooth regions that aresimple to reconstruct, we propose a sampling strategy to select challengingtraining patches during the HDR fine-tuning stage. We demonstrate throughexperimental results that our approach can reconstruct visually pleasing HDRresults, better than the current state of the art on a wide range of scenes.