No Reference Quality Assessment for Screen Content Images With Both Local and Global Feature Representation

Area of Science:

• Computer Vision
• Image Processing
• Human Visual Perception

Background:

• Screen content images (SCIs) present unique challenges for quality assessment.
• No-reference (NR) quality assessment is crucial due to the absence of original, distortion-free images.
• Human visual system (HVS) perception relies on luminance and texture information.

Purpose of the Study:

• To propose a novel no-reference quality assessment method for SCIs.
• To incorporate statistical luminance and texture features for robust quality prediction.
• To develop a method inspired by HVS perceptual properties.

Main Methods:

• Calculating a luminance map via local normalization.
• Extracting statistical luminance features globally.
• Utilizing high-order derivatives and local binary patterns on gradient maps for texture analysis.
• Employing support vector regression to map features to subjective ratings.

Main Results:

• The proposed NRLT method demonstrates superior performance in predicting SCI visual quality.
• NRLT achieves better results compared to existing NR methods.
• The method even surpasses some full-reference quality assessment techniques.

Conclusions:

• The NRLT method offers an effective approach for no-reference quality assessment of SCIs.
• Integrating statistical luminance and texture features enhances prediction accuracy.
• The method shows promise for real-world applications requiring image quality evaluation.