An Efficient Method for No-Reference Video Quality Assessment.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• The proliferation of user-generated videos necessitates robust methods for assessing video quality without reference.
• Existing No-Reference Video Quality Assessment (NR-VQA) methods often face challenges with natural distortions and computational efficiency.

Purpose of the Study:

• To develop an effective and computationally efficient NR-VQA method for consumer-produced videos.
• To improve the generalization capability of NR-VQA models across different video databases.

Main Methods:

• A novel frame sampling module selects key frames for quality assessment.
• Two lightweight Convolutional Neural Networks (CNNs) encode video frame attributes and semantic content.
• A Support Vector Regressor (SVR) estimates the overall video quality score.

Main Results:

• The proposed NR-VQA method achieves performance comparable to state-of-the-art methods on benchmark datasets (CVD2014, KoNViD-1k, LIVE-Qualcomm, LIVE-VQC).
• The method demonstrates significantly lower computational cost compared to existing approaches.
• Superior generalization performance is observed in cross-database testing scenarios.

Conclusions:

• The developed NR-VQA method offers an efficient and accurate solution for evaluating video quality.
• The frame sampling strategy and lightweight CNNs contribute to reduced computational load without sacrificing performance.
• The method shows promise for real-world applications involving large-scale video quality assessment.