EvoIQA - Explaining Image Distortions with Evolved White-Box Logic

A research team from academia has introduced EvoIQA, a novel approach to Image Quality Assessment (IQA) that bridges the gap between rigid mathematical models and opaque deep learning systems. Using Genetic Programming—an evolutionary algorithm technique—the framework automatically generates human-readable mathematical formulas that assess image quality. Unlike traditional "black-box" neural networks, EvoIQA produces explicit equations that map structural, chromatic, and information-theoretic degradations, making every prediction fully interpretable. The system builds on components from established metrics like VSI, VIF, FSIM, and HaarPSI to create evolved models that explain exactly why an image appears distorted.

In testing, EvoIQA's evolved models demonstrated strong alignment with human visual preferences and achieved performance parity with complex, state-of-the-art deep learning architectures such as DB-CNN. This represents a significant breakthrough in explainable AI for computer vision, proving that high accuracy no longer requires sacrificing interpretability. The research, detailed in an 11-page arXiv paper, shows that symbolic regression can compete with neural networks on perceptual tasks while providing the transparency needed for critical applications in medical imaging, autonomous systems, and content moderation where understanding AI decisions is essential.