Blind image quality assessment via probabilistic latent semantic analysis.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Blind image quality assessment (BIQA) is crucial for evaluating image fidelity without reference images.
• Existing BIQA methods often require extensive training data and complex models.
• A need exists for unsupervised, training-free BIQA methods.

Purpose of the Study:

• To develop a novel, unsupervised, and training-free blind image quality assessment method.
• To leverage latent characteristics and probabilistic topic modeling for image quality evaluation.
• To establish a BIQA approach that aligns with human subjective judgments.

Main Methods:

• Utilized probabilistic latent semantic analysis and a visual word dictionary.
• Extracted four distortion-affected features: block-based local histogram, local mean, contrast mean, and contrast variance.
• Quantified image quality by measuring topic frequency discrepancy between distorted and pristine images.

Main Results:

• The proposed method demonstrates high correlation with human subjective quality assessments across diverse distortions.
• Experimental results on four open databases validate the method's effectiveness.
• The approach successfully identifies latent characteristics indicative of image distortion.

Conclusions:

• The developed unsupervised, training-free BIQA method offers a promising alternative to traditional approaches.
• The technique effectively utilizes topic modeling on visual features for robust quality assessment.
• This method provides a reliable and efficient way to evaluate image quality objectively.