Burst Denoising of Dark Images

  • 2020-03-17 17:17:36
  • Ahmet Serdar Karadeniz, Erkut Erdem, Aykut Erdem
  • 71

Abstract

Capturing images under extremely low-light conditions poses significantchallenges for the standard camera pipeline. Images become too dark and toonoisy, which makes traditional image enhancement techniques almost impossibleto apply. Very recently, researchers have shown promising results usinglearning based approaches. Motivated by these ideas, in this paper, we proposea deep learning framework for obtaining clean and colorful RGB images fromextremely dark raw images. The backbone of our framework is a novelcoarse-to-fine network architecture that generates high-quality outputs in aprogressive manner. The coarse network predicts a low-resolution, denoised rawimage, which is then fed to the fine network to recover fine-scale details andrealistic textures. To further reduce noise and improve color accuracy, weextend this network to a permutation invariant structure so that it takes aburst of low-light images as input and merges information from multiple imagesat the feature-level. Our experiments demonstrate that the proposed approachleads to perceptually more pleasing results than state-of-the-art methods byproducing much sharper and higher quality images.

 

Quick Read (beta)

loading the full paper ...