A Perceptually Optimized and Self-Calibrated Tone Mapping Operator.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• High dynamic range (HDR) photography is increasingly popular, necessitating efficient tone mapping operators (TMOs) for dynamic range compression.
• Existing TMOs face challenges in balancing visual quality and computational speed.
• Human visual system (HVS) principles offer insights for developing perceptually optimized image processing techniques.

Purpose of the Study:

• To develop a novel, self-calibrated, and perceptually optimized two-stage neural network-based TMO for HDR image tone mapping.
• To improve the visual quality of tone-mapped images by aligning with human perception.
• To achieve high performance comparable to fast local TMOs.

Main Methods:

• A two-stage neural network approach was employed, inspired by human visual system physiology.
• Stage one decomposes HDR images into a normalized Laplacian pyramid and uses lightweight neural networks to estimate the LDR image's Laplacian pyramid, optimized via a perceptual metric.
• Stage two generates a pseudo-multi-exposure image stack and uses a fusion network to merge them, maximizing a structural similarity index variant.

Main Results:

• The proposed method consistently produces images with superior visual quality compared to existing TMOs.
• The developed TMO ranks among the fastest local TMOs in terms of computational speed.
• Experimental results validate the perceptual relevance and effectiveness of the optimization metrics used.

Conclusions:

• The two-stage neural network TMO offers a significant advancement in HDR image tone mapping.
• The self-calibrated and perceptually optimized approach effectively enhances image detail and color saturation.
• This method provides a promising solution for real-time HDR image processing applications.