Tone Mapping of HDR Images via Meta-Guided Bayesian Optimization and Virtual Diffraction Modeling.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• High-dynamic-range (HDR) image tone mapping aims to represent scenes with extreme lighting variations.
• Existing methods often struggle to balance structural fidelity with perceptual naturalness.
• Human visual perception of luminance is nonlinear, requiring specialized modeling.

Purpose of the Study:

• To develop a novel tone-mapping framework that enhances both perceptual fidelity and visual naturalness in HDR images.
• To improve the efficiency and robustness of HDR tone mapping through meta-learning and Bayesian optimization.
• To address the cold-start problem and accelerate convergence in parameter tuning.

Main Methods:

• Virtual diffraction formalized as a frequency-domain operator for HDR image reconstruction.
• Application of Stevens' power law to model human nonlinear luminance perception.
• Bayesian optimization for adaptive parameter tuning to optimize the Tone Mapping Quality Index (TMQI).
• A task-distribution-oriented meta-learning framework for rapid parameter prediction and generalization.

Main Results:

• The proposed method achieves superior performance compared to state-of-the-art tone-mapping algorithms on benchmark datasets.
• An average improvement of up to 27% in visual naturalness was observed.
• Meta-learning-guided Bayesian optimization demonstrated a two- to five-fold increase in convergence speed.
• The framework consistently dominates the Pareto frontier for computational time versus performance.

Conclusions:

• The novel framework effectively balances perceptual fidelity and visual naturalness in HDR image tone mapping.
• Meta-learning significantly enhances robustness and mitigates the cold-start problem, accelerating optimization.
• The approach offers high-quality results with efficient convergence and low computational cost.