Structural Similarity in Deep Features: Unified Image Quality Assessment Robust to Geometrically Disparate Reference.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional Image Quality Assessment (IQA) methods often assume perfect pixel alignment, limiting their effectiveness in real-world scenarios with geometric distortions.
• Existing approaches for Geometrically-Disparate-Reference IQA (GDR-IQA) are typically task-specific or rely on assumptions about distortion magnitude.

Purpose of the Study:

• To propose a unified, non-training-based framework for Image Quality Assessment (IQA) that addresses both Aligned-Reference IQA (AR-IQA) and Geometrically-Disparate-Reference IQA (GDR-IQA).
• To develop a method that assesses structural similarity of deep features efficiently and robustly, without requiring task-specific designs.

Main Methods:

• Introduced Deep Structural Similarity (DeepSSIM), a non-training-based approach evaluating structural similarity in deep feature space.
• Incorporated an attention calibration strategy to mitigate attention deviation issues.
• Applied the method to both AR-IQA and GDR-IQA benchmarks.

Main Results:

• Achieved state-of-the-art performance on established AR-IQA datasets.
• Demonstrated strong robustness across various GDR-IQA test cases, outperforming existing methods.
• Showcased DeepSSIM's effectiveness as an optimization tool for image enhancement tasks like super-resolution and restoration.

Conclusions:

• DeepSSIM provides a versatile and high-performing solution for IQA problems, unifying the assessment of aligned and geometrically distorted images.
• The method's generalizability extends beyond quality assessment, proving useful in optimizing image restoration and enhancement model training.