Learning No-Reference Quality Assessment of Multiply and Singly Distorted Images with Big Data.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• No-reference (NR) image quality assessment (IQA) traditionally focused on limited distortion types.
• Existing algorithms like MUSIQUE struggle with diverse, real-world image degradations.
• Images encounter multiple distortions (noise, blur, compression, contrast changes) during processing.

Purpose of the Study:

• To extend the MUSIQUE algorithm for comprehensive NR IQA.
• To develop a robust algorithm for assessing images with five distortion types and their combinations.
• To improve the accuracy and scope of blind image quality prediction.

Main Methods:

• Introduced MUSIQUE-II, building on MUSIQUE's framework with advanced models and features.
• Employed a three-layer classification model to identify 19 distortion types.
• Extracted 14 contrast features and used a multi-layer probability-weighting rule for parameter estimation.
• Implemented a most-apparent-distortion strategy to combine quality scores adaptively.

Main Results:

• MUSIQUE-II demonstrated significant improvements in quality prediction performance over its predecessor.
• The algorithm achieved highly competitive results compared to state-of-the-art full-reference (FR) and NR IQA methods.
• Experimental validation on multiple databases confirmed its effectiveness on singly and multiply-distorted images.

Conclusions:

• MUSIQUE-II offers a more versatile and accurate solution for NR IQA in practical scenarios.
• The enhanced feature set and classification framework enable better handling of diverse image distortions.
• This work advances the field of blind image quality assessment for complex image degradation.