No-Reference Quality Assessment of Authentically Distorted Images Based on Local and Global Features

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Computing

Background:

• Digital imaging advancements necessitate robust image quality assessment (IQA).
• Lack of distortion-free references in practical applications drives demand for no-reference IQA (NR-IQA).
• Authentic image distortions require specialized evaluation methods.

Purpose of the Study:

• To introduce a novel NR-IQA algorithm for objectively evaluating authentically distorted images.
• To characterize diverse authentic distortions using comprehensive feature vectors.
• To enhance the performance of NR-IQA through a combination of local and global features.

Main Methods:

• Utilized a wide array of local and global feature vectors to represent image distortions.
• Employed statistics of established local feature descriptors (SURF, FAST, BRISK, KAZE) for NR-IQA.
• Introduced additional features to improve the efficacy of local feature-based analysis.

Main Results:

• The proposed algorithm was evaluated against 12 state-of-the-art methods.
• Performance was assessed on benchmark datasets (CLIVE, KonIQ-10k, SPAQ) with authentic distortions.
• The novel algorithm demonstrated significantly superior performance in predicting human perceptual quality ratings.

Conclusions:

• The developed NR-IQA algorithm offers a robust solution for evaluating authentically distorted images.
• The method's reliance on diverse feature characterization contributes to its high accuracy.
• This work advances objective image quality assessment in scenarios lacking ground truth references.