Automatic Intermediate Generation With Deep Reinforcement Learning for Robust Two-Exposure Image Fusion.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• High dynamic range (HDR) image generation from low dynamic range (LDR) images is crucial for applications with limited hardware.
• Current methods struggle with insufficient input images or poorly exposed content, leading to suboptimal HDR results.

Purpose of the Study:

• To develop a novel deep reinforcement learning framework for two-exposure image fusion to generate high dynamic range (HDR) images.
• To address limitations in existing HDR generation techniques, particularly when dealing with limited well-exposed input images.

Main Methods:

• Modeling HDR image generation as a deep reinforcement learning problem for two-exposure fusion.
• Learning an online compensating representation to fuse with LDR inputs.
• Creating and utilizing a new two-exposure dataset with reference HDR images for training and evaluation.

Main Results:

• The proposed reinforcement learning method significantly outperforms competing methods in HDR image generation, even with limited well-exposed content.
• Experimental results on a no-reference multiexposure dataset confirm the model's generality and effectiveness.
• Demonstrated superior performance across various challenging scenarios.

Conclusions:

• This work presents the first reinforcement-learning-based framework for online compensating representation in two-exposure image fusion for HDR generation.
• The developed approach offers a robust solution for HDR image generation, overcoming limitations of traditional methods and demonstrating broad applicability.