No-Reference Quality Assessment for Screen Content Images Using Visual Edge Model and AdaBoosting Neural Network

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Assessing perceptual quality of screen content images (SCIs) is crucial for user experience.
• Existing no-reference metrics often struggle with the unique characteristics of SCIs.
• The human visual system's sensitivity to edge information is a key factor in perceived image quality.

Purpose of the Study:

• To propose a novel no-reference metric for assessing the perceptual quality of SCIs.
• To leverage human visual principles and advanced machine learning for accurate quality assessment.
• To develop a robust method that aligns with human perception.

Main Methods:

• Utilized the difference of Gaussians (DOG) model to extract multi-scale edge maps (contour and edge information).
• Employed L-moments distribution estimation for locally normalized edge maps to derive edge features.
• Developed an AdaBoosting back-propagation neural network (ABPNN) with a deep architecture to map features to perceptual quality scores.

Main Results:

• The proposed metric demonstrated highly competitive performance in assessing SCI perceptual quality.
• The method showed significant consistency with the human visual system (HVS).
• The deep architecture of the ABPNN contributed to good generalization ability.

Conclusions:

• The developed no-reference metric effectively assesses SCI perceptual quality by integrating human visual edge models and an ABPNN.
• The approach provides a robust and accurate method for image quality assessment, aligning well with human perception.
• This work offers a promising solution for evaluating SCIs without requiring original reference images.