Full-Reference Image Quality Expression via Genetic Programming

Area of Science:

• Computer Vision
• Image Processing
• Computational Intelligence

Background:

• Full-reference image quality measures are essential for digital data management tasks like retrieval, compression, and copyright protection.
• Existing measures like the Structural Similarity Index Measure (SSIM) are effective but can be improved.
• Developing accurate image quality assessment (IQA) methods that mimic human perception is a key challenge.

Purpose of the Study:

• To develop a framework for automatically generating SSIM-like image quality measures using genetic programming.
• To optimize these measures for better correlation with human mean opinion scores (MOS).
• To explore the trade-off between measure complexity and performance.

Main Methods:

• Utilized genetic programming to evolve new image quality assessment algorithms.
• Defined terminal sets based on structural similarity concepts at various abstraction levels.
• Implemented a two-stage genetic optimization process incorporating hoist mutation to manage solution complexity.
• Validated optimized measures across multiple datasets to ensure generalizability.

Main Results:

• The proposed framework successfully generated novel SSIM-like image quality measures.
• Optimized measures demonstrated superior performance compared to existing SSIM variants in predicting human MOS.
• Performance was evaluated using cross-dataset validation, ensuring robustness.
• Tuning measures to specific datasets yielded competitive or superior results compared to more complex IQA methods.

Conclusions:

• Genetic programming provides an effective approach for developing advanced image quality assessment metrics.
• The developed SSIM-like measures offer a promising balance of simplicity and high performance.
• These findings contribute to more accurate and efficient digital image management systems.