Multi-Dimensional Feature Fusion Network for No-Reference Quality Assessment of In-the-Wild Videos

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Video quality assessment (VQA) is crucial, but in-the-wild video quality prediction without reference faces challenges from hybrid distortions and content motion.
• Existing no-reference VQA (NR-VQA) methods struggle with dynamic variations and complex distortions.

Purpose of the Study:

• To develop an advanced NR-VQA method that effectively integrates static and dynamic visual information for accurate quality prediction.
• To improve the perception of video quality in challenging 'in-the-wild' scenarios.

Main Methods:

• Utilized multi-dimensional convolutional networks for extracting fused static-dynamic features from video clips.
• Implemented alignment and a temporal memory module with recurrent neural networks and fully connected branches for time-series feature association.
• Developed a parametric adaptive network structure to mimic human visual perception for final score generation.

Main Results:

• The proposed NR-VQA method demonstrated superior performance on mixed datasets (CVD2014, KoNViD-1k, LIVE-Qualcomm, LIVE-VQC).
• Achieved competitive results on individual datasets, validating its generalization ability.
• Outperformed existing state-of-the-art NR-VQA techniques.

Conclusions:

• The novel NR-VQA method effectively addresses challenges posed by hybrid distortions and dynamic content.
• The integration of static-dynamic features and human visual habit simulation leads to robust video quality prediction.
• The method shows significant potential for real-world applications requiring accurate video quality assessment.