No-Reference Video Quality Assessment Using the Temporal Statistics of Global and Local Image Features.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Digital video quality degrades during acquisition, storage, and transmission, impacting computer vision tasks.
• No-reference video quality assessment (NR-VQA) algorithms are crucial for evaluating video quality without a pristine reference.
• Existing NR-VQA methods face challenges in accurately assessing diverse video distortions.

Purpose of the Study:

• To propose a novel NR-VQA algorithm that enhances video quality assessment accuracy.
• To develop an integrated architecture combining local and global feature statistics with ensemble learning.
• To effectively characterize a wide range of video distortions using temporal feature statistics.

Main Methods:

• Developed a novel NR-VQA algorithm integrating temporal statistics of local and global image features.
• Employed an ensemble learning framework within a single architecture for robust quality prediction.
• Utilized a broad spectrum of statistical measures to capture various video distortions.

Main Results:

• The proposed NR-VQA algorithm achieved significantly improved results on benchmark datasets (KoNViD-1k, LIVE VQC).
• Experimental results demonstrated superior performance compared to 14 other well-known NR-VQA algorithms.
• The method showed substantial progress over recent NR-VQA approaches in accuracy and effectiveness.

Conclusions:

• The proposed fusion of temporal statistics and ensemble learning offers a powerful approach for NR-VQA.
• The algorithm effectively addresses the challenge of assessing video quality without a reference.
• This work advances the field of video quality assessment, benefiting computer vision applications.