Personalized Exposure Control Using Adaptive Metering and Reinforcement Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Computational Photography

Background:

• Real-time mobile camera exposure control is challenging.
• Existing methods often lack personalization and struggle with optimizing multiple parameters simultaneously.

Purpose of the Study:

• To develop a personalizable reinforcement learning approach for real-time mobile camera exposure control.
• To optimize the trade-off between image quality, convergence speed, and temporal stability.

Main Methods:

• Utilized a Markov Decision Process (MDP) framework.
• Modeled exposure prediction using a fully convolutional neural network trained via Gaussian policy gradient.
• Incorporated an adaptive metering module linking scene semantics with exposure settings.

Main Results:

• The system demonstrates stable real-time performance.
• Achieved improved visual quality compared to native camera controls.
• Showcased personalization capabilities for users and devices.

Conclusions:

• The proposed reinforcement learning approach offers effective and personalizable real-time exposure control for mobile cameras.
• The adaptive metering module enhances generalization beyond conventional techniques.
• The end-to-end trained neural network successfully associates scene semantics with optimal exposure values.