Dynamic Receptive Field Generation for Full-Reference Image Quality Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Computing

Background:

• Current full-reference image quality assessment (FR-IQA) methods often lack specificity to distortion types.
• Image distortion perception is nonlinearly dependent on the specific type of distortion present.

Purpose of the Study:

• To propose a novel FR-IQA framework that dynamically adapts to different image distortion types.
• To enhance prediction accuracy in image quality assessment by considering distortion-specific characteristics.

Main Methods:

• Introduced a dynamic receptive field generation based image quality assessor (DRF-IQA) framework.
• Implemented a two-stream approach: dynamic error representation and visual sensitivity-based quality pooling.
• Utilized a convolutional neural network (CNN) to generate dynamic receptive fields and spatial error maps, weighted by a visual sensitivity map.

Main Results:

• The DRF-IQA model demonstrated state-of-the-art prediction accuracy across multiple open image quality assessment databases.
• The dynamic generation of receptive fields proved effective in capturing distortion-specific perceptual information.
• The visual sensitivity-based pooling enhanced the correlation between predicted and actual image quality.

Conclusions:

• The proposed DRF-IQA framework offers a significant advancement in full-reference image quality assessment.
• Dynamically adapting to distortion types leads to more accurate and perceptually relevant image quality predictions.
• This approach provides a more robust solution for evaluating image quality in diverse scenarios.