Perceptual Texture Similarity Estimation: An Evaluation of Computational Features.

Area of Science:

• Computer Vision
• Material Science
• Human-Computer Interaction

Background:

• Accurate texture similarity estimation is crucial for material recognition.
• Existing texture feature sets often fail to align with human perceptual judgments.
• Human perception of texture involves complex, long-range interactions.

Purpose of the Study:

• To comprehensively evaluate 51 texture feature sets against human perceptual similarity data.
• To investigate the effectiveness of feature combination methods.
• To explore the potential of Convolutional Neural Networks (CNNs) for improved texture similarity estimation.

Main Methods:

• Evaluation of 51 texture feature sets using fine-grained human perceptual similarity ground-truth.
• Comparison of individual feature performance against human agreement rates.
• Application of Random Forest for feature combination.
• Statistical modeling to analyze feature capabilities in capturing relationships.
• Psychophysical experiments to understand human texture perception cues.
• Assessment of pre-trained Convolutional Neural Networks (CNNs) for texture similarity tasks.

Main Results:

• Most individual texture features poorly estimate perceptual similarity compared to human agreement.
• Combining features using Random Forests significantly enhances performance.
• Few traditional features capture essential long-range, aperiodic relationships.
• Human perception relies on long-range interactions for texture similarity judgment.
• CNNs, particularly their deeper layers and extensive filters, demonstrate a significant advantage.

Conclusions:

• Traditional texture features are insufficient for accurately modeling human perceptual similarity.
• Feature combination and advanced deep learning models like CNNs are necessary for robust texture similarity estimation.
• CNNs excel due to their ability to learn features over large spatial extents, mirroring human perceptual strategies.