Blind Quality Estimation by Disentangling Perceptual and Noisy Features in High Dynamic Range Images.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• High dynamic range (HDR) imaging offers more realistic visual content, aligning with human perception.
• Assessing HDR image quality without a reference is difficult due to the cost of HDR displays.
• Existing research on no-reference image quality assessment (NR-IQA) for HDR data is limited.

Purpose of the Study:

• To develop a novel NR-IQA model for distorted HDR images.
• To address the challenge of evaluating HDR visual quality without a reference.
• To leverage convolutional neural networks for artifact detection and quality prediction.

Main Methods:

• A convolutional neural network (CNN)-based NR-IQA model was proposed for HDR images.
• The model separately measures error and perceptual masking effects.
• Perceptual masking effects were learned from annotated HDR image datasets during training.

Main Results:

• The proposed NR-IQA model effectively detects visual artifacts in distorted HDR images.
• The model considers perceptual masking effects for more accurate quality prediction.
• Experimental results show the model's performance is comparable to state-of-the-art full-reference IQA methods.

Conclusions:

• The developed NR-IQA model provides accurate quality assessment for HDR images without requiring a reference.
• This approach offers a viable solution for evaluating HDR image quality, overcoming previous limitations.
• The model's ability to predict perceived quality demonstrates its practical applicability in HDR imaging systems.