Abstract
Images captured with unreasonable exposures will greatly reduce visual quality. Exposure problems can be categorized as follows: (i) over-exposure, i.e., bright and losing image regions caused by too-long exposure; (ii) under-exposure, i.e., dark and drowned-in noises caused by too-short exposure. Most prior works only handle over- or under-exposure problems on sRGB domain and ignore prior knowledge of channel information. In this paper, we propose Deep Guided network for exposure correction on RAW domain with Knowledge Distillation (denoted as DGKD), solving two problems together. Firstly, according to color sensitivity, we employ blue/red channel and green channel as guidance information for over- and under-exposure correction, respectively. Secondly, to handle two varying problems in a unified network, we first train the over- and under-exposure correction networks individually and then distill knowledge into one deep guided network. The experimental results show that the proposed method outperforms the state-of-the-art methods under both quantitative metrics and visual quality. Specifically, the proposed method attained a peak signal-to-noise ratio of 24.653 dB and a structural similarity index of 0.8182 on the collected RAW image exposure correction dataset.