A No-Reference Texture Regularity Metric Based on Visual Saliency

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Quality Assessment

Background:

• Assessing perceived texture regularity is crucial for image processing.
• Existing methods often require reference images or lack perceptual accuracy.
• Understanding human visual attention is key to developing better metrics.

Purpose of the Study:

• To propose a no-reference perceptual metric for quantifying texture regularity.
• To develop a ground-truth eye-tracking database for texture perception.
• To evaluate and validate the proposed metric against human subjective scores.

Main Methods:

• Developed a no-reference metric based on visual attention (VA) similarity and periodic spatial distribution.
• Created a ground-truth eye-tracking database for texture perception.
• Utilized saliency maps from the best-performing VA model to compute the metric.

Main Results:

• The proposed metric shows a strong correlation with subjective mean opinion scores for perceived texture regularity.
• The metric effectively quantifies the degree of perceived regularity in textures.
• Validated the performance of popular VA models using the generated eye-tracking database.

Conclusions:

• The novel texture regularity metric accurately reflects human perception.
• This metric can enhance image processing applications such as texture synthesis, compression, and retrieval.
• The developed eye-tracking database serves as a valuable resource for VA model evaluation.