Objective Video Quality Assessment Combining Transfer Learning With CNN.

Area of Science:

• Computer Vision
• Signal Processing

Background:

• Video quality assessment (VQA) is crucial for video compression, transmission, and storage.
• Existing VQA methods are hindered by small, imbalanced databases and limited feature representations of distorted videos.

Purpose of the Study:

• To propose a novel full-reference (FR) VQA metric that overcomes limitations of current VQA databases and feature extraction methods.
• To enhance the development of VQA methods through transfer learning and deep neural networks.

Main Methods:

• Utilized a feature-based transfer learning framework to enrich distorted video samples.
• Employed a six-layer convolutional neural network (CNN) for high-level spatiotemporal feature extraction from image blocks (IBs) and video blocks (VBs).
• Incorporated preprocessing and postprocessing steps to mitigate inaccuracies from predicted labels by classic FR metrics, using saliency maps and entropy for pooling block-level scores.

Main Results:

• The proposed FR-VQA metric demonstrated high expandability through iterative train-test procedures and cross-database testing.
• Achieved performance comparable to state-of-the-art VQA metrics on widely used databases with diverse compression distortions.

Conclusions:

• The developed VQA metric effectively addresses challenges of limited data and feature representation in distorted videos.
• Transfer learning and CNN integration provide a robust and scalable solution for accurate video quality assessment.