Image Quality Assessment: Unifying Structure and Texture Similarity.

Area of Science:

• Computer Vision
• Image Processing
• Perceptual Science

Background:

• Traditional objective image quality measures struggle with texture resampling, unlike human observers.
• Existing models are overly sensitive to spatial alterations in textured regions.

Purpose of the Study:

• Develop a full-reference image quality model with explicit tolerance to texture resampling.
• Create a model that aligns better with human perceptual judgments of image quality.

Main Methods:

• Utilized a convolutional neural network to create multi-scale overcomplete image representations.
• Developed a novel image quality metric combining "texture similarity" and "structure similarity" based on feature map analysis.
• Optimized model parameters using human quality ratings and minimizing subimage distances within textures.

Main Results:

• The proposed method, DISTS, effectively captures texture appearance and synthesizes diverse texture patterns.
• DISTS demonstrates strong correlation with human perceptual scores across conventional and texture-specific image databases.
• Achieved competitive performance in texture classification and retrieval tasks.

Conclusions:

• The developed image quality model offers improved robustness to geometric transformations without specialized training.
• This approach provides a more perceptually relevant objective measure for image quality assessment, especially for textured images.