Personalized Exposure Control Using Adaptive Metering and Reinforcement Learning

  • 2018-08-05 08:05:27
  • Huan Yang, Baoyuan Wang, Noranart Vesdapunt, Minyi Guo, Sing Bing Kang
  • 1

Abstract

We propose a reinforcement learning approach for real-time exposure controlof a mobile camera that is personalizable. Our approach is based on MarkovDecision Process (MDP). In the camera viewfinder or live preview mode, giventhe current frame, our system predicts the change in exposure so as to optimizethe trade-off among image quality, fast convergence, and minimal temporaloscillation. We model the exposure prediction function as a fully convolutionalneural network that can be trained through Gaussian policy gradient in anend-to-end fashion. As a result, our system can associate scene semantics withexposure values; it can also be extended to personalize the exposureadjustments for a user and device. We improve the learning performance byincorporating an adaptive metering module that links semantics with exposure.This adaptive metering module generalizes the conventional spot or matrixmetering techniques. We validate our system using the MIT FiveK and our owndatasets captured using iPhone 7 and Google Pixel. Experimental results showthat our system exhibits stable real-time behavior while improving visualquality compared to what is achieved through native camera control.

 

Quick Read (beta)

loading the full paper ...