Blind Deep S3D Image Quality Evaluation via Local to Global Feature Aggregation.

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Traditional no-reference stereoscopic 3D image quality assessment (NR-S3D-IQA) relies on hand-crafted features, limiting accuracy.
• Existing NR-S3D-IQA methods struggle to match the performance of full-reference (FR) metrics due to feature optimization challenges.

Purpose of the Study:

• To develop a novel deep learning scheme for NR-S3D-IQA that overcomes limitations of conventional approaches.
• To enhance the prediction accuracy of S3D image quality scores without requiring depth estimation.

Main Methods:

• A deep convolutional neural network (CNN) trained via two-step regression for local to global feature aggregation.
• Utilized a full-reference S3D-IQA metric for initial training data approximation, followed by iterative updates using subjective mean opinion scores (MOS).
• Implemented a local patch-based CNN approach, aggregating features through a dedicated layer without explicit depth estimation.

Main Results:

• The proposed deep learning scheme significantly improves NR-S3D-IQA performance compared to previous algorithms.
• Achieved accuracy competitive with FR-S3D-IQA metrics, demonstrating approximately 91% correlation with MOS.
• The method effectively predicts S3D image quality scores using aggregated local and global features.

Conclusions:

• The novel deep learning approach offers a superior method for NR-S3D-IQA.
• The local to global feature aggregation strategy enhances prediction accuracy and robustness.
• This method provides a valuable tool for evaluating stereoscopic 3D image quality efficiently and accurately.