Facebook
From Baby Mosquito, 3 Years ago, written in Plain Text.
Embed
Download Paste or View Raw
Hits: 118
  1.  
  2. While previous works have focused on specific aspects of low- light imaging (such as denoising (Dabov et al. 2007; Mildenhall et al. 2018)), there have been relatively few works that describe photographic systems that address more than one of the above re- quirements. The main strategies that have been used by previous systems are burst imaging (Hasinoff et al. 2016), and end-to-end trained convolutional neural networks (CNNs) (Chen et al. 2018). Burst imaging systems capture, align, and merge multiple frames to generate a temporally-denoised image. This image is then further processed using a series of operators including additional spatial denoising, white balancing, and tone mapping. CNN-based sys- tems attempt to perform as much of this pipeline as possible using a deep neural network with a single raw frame as its input. The network learns how to perform all of the image adjustment opera- tors. CNN-based solutions often require significant computational resources (memory and time) and it is challenging to optimize their performance so that they could run on mobile devices. Additionally, end-to-end learning based systems don’t provide a way to design and tune an imaging pipeline; they can only imitate the pipeline they were trained on.