Chen Xiangyu ; Liu Yihao ; Zhang Zhengwen ; Qiao Yu ; Dong Chao

Abstract

Most consumer-grade digital cameras can only capture a limited range ofluminance in real-world scenes due to sensor constraints. Besides, noise andquantization errors are often introduced in the imaging process. In order toobtain high dynamic range (HDR) images with excellent visual quality, the mostcommon solution is to combine multiple images with different exposures.However, it is not always feasible to obtain multiple images of the same sceneand most HDR reconstruction methods ignore the noise and quantization loss. Inthis work, we propose a novel learning-based approach using a spatially dynamicencoder-decoder network, HDRUNet, to learn an end-to-end mapping for singleimage HDR reconstruction with denoising and dequantization. The networkconsists of a UNet-style base network to make full use of the hierarchicalmulti-scale information, a condition network to perform pattern-specificmodulation and a weighting network for selectively retaining information.Moreover, we propose a Tanh_L1 loss function to balance the impact ofover-exposed values and well-exposed values on the network learning. Our methodachieves the state-of-the-art performance in quantitative comparisons andvisual quality. The proposed HDRUNet model won the second place in the singleframe track of NITRE2021 High Dynamic Range Challenge.