No-Reference Image Quality Assessment with Multi-Scale Orderless Pooling of Deep Features.

Area of Science:

• Computer Vision
• Multimedia Signal Processing
• Machine Learning

Background:

• No-reference image quality assessment (NR-IQA) is crucial for evaluating digital images without pristine references.
• Images suffer various distortions during processing, transmission, and storage.
• Existing NR-IQA methods require improvement in feature extraction effectiveness.

Purpose of the Study:

• To propose a novel convolutional neural network (CNN) architecture for enhanced NR-IQA.
• To improve the accuracy of perceptual quality assessment for distorted images.
• To develop a method that effectively extracts deep features at multiple scales.

Main Methods:

• A novel CNN architecture for NR-IQA is proposed.
• Deep features are extracted from local image patches at multiple scales.
• Gaussian process regressors are trained to map extracted features to perceptual quality scores.

Main Results:

• The proposed algorithm demonstrates superior performance compared to state-of-the-art methods.
• Experiments were conducted on three large benchmark datasets: LIVE In the Wild, KonIQ-10k, and SPAQ.
• The method shows favorable results on datasets with authentic distortions.

Conclusions:

• The novel multi-scale deep feature extraction architecture significantly enhances NR-IQA performance.
• The proposed method offers a robust solution for evaluating perceptual image quality without reference images.
• This approach advances the field of automated image quality assessment.