The Importance of Phase to Texture Discrimination and Similarity

Area of Science:

• Computer Vision
• Human Perception
• Image Processing

Background:

• Texture discrimination and similarity estimation are crucial in image analysis.
• The relative importance of phase and magnitude spectra in visual processing is not fully understood.
• Computational models often differ from human perception in feature utilization.

Purpose of the Study:

• To investigate the role of phase information in human texture perception.
• To compare human performance with computational feature sets in texture tasks.
• To improve computational models for texture similarity estimation.

Main Methods:

• Conducted two psychophysical experiments to assess human reliance on phase and magnitude.
• Evaluated 51 computational feature sets on texture discrimination and similarity tasks.
• Investigated feature set performance on phase-only images and incorporated Convolutional Neural Network (CNN) features.

Main Results:

• Psychophysical experiments revealed human preference for phase information in texture perception.
• Computational feature sets predominantly utilized magnitude data, contrasting with human perception.
• A novel feature fusion scheme combining CNN features significantly improved texture similarity estimation accuracy.

Conclusions:

• Phase information is critical for human texture perception, particularly for similarity estimation.
• Current computational models may not effectively capture the importance of phase for texture analysis.
• Feature fusion strategies, especially those incorporating CNNs, can enhance computational models to better align with human visual perception of textures.