Kernel Prediction Network for Detail-Preserving High Dynamic Range Imaging

Abstract

Generating a high dynamic range (HDR) image from multiple exposure images is challenging in the presence of significant motions, which usually causes ghosting artifacts. To alleviate this problem, previous methods explicitly align the input images before merging the controlled exposure images. Although recent works try to learn the HDR imaging process using a convolutional neural network (CNN), they still suffer from ghosting or blurring artifacts and missing details in extremely under/overexposed areas. In this paper, we propose an end-to-end framework for detail-preserving HDR imaging of dynamic scenes. Our method employs a kernel prediction network and produces per-pixel kernels to fully utilize every pixel and its neighborhood in input images for the successful alignment. After applying the kernels to the input images, we generate a final HDR image using a simple merging network. The proposed framework is an end-to-end trainable method without any preprocessing, which not only avoids ghosting or blurring artifacts but also hallucinates fine details effectively. We demonstrate that our method provides comparable results to the state-of-the-art methods regarding qualitative and quantitative evaluations.