Effective and efficient subjective testing of texture similarity metrics

Area of Science:

• Computer Vision
• Image Processing
• Human-Computer Interaction

Background:

• Objective texture similarity metrics are crucial for image analysis but require validation against human perception.
• Current subjective testing methods for texture similarity can be inefficient and difficult to scale.
• Understanding human visual perception is key to developing reliable texture similarity metrics.

Purpose of the Study:

• To define performance requirements and testing procedures for objective texture similarity metrics.
• To introduce and validate a novel subjective testing procedure, ViSiProG, for evaluating texture similarity.
• To assess the performance of texture similarity metrics across different perceptual domains.

Main Methods:

• Defined three operating domains for evaluating texture similarity metrics: identical textures, top of the similarity scale, and distinguishing similar/dissimilar textures.
• Proposed ViSiProG (Visual Similarity by Progressive Grouping), a procedure organizing texture databases into visually similar clusters using visual blending.
• Conducted subjective experiments using ViSiProG to collect data on texture similarity judgments.

Main Results:

• ViSiProG simplifies the labeling of image pairs as similar or dissimilar, enhancing data collection efficiency.
• Experimental results demonstrated the effectiveness of ViSiProG in gathering subjective data for large texture databases.
• Comparisons with existing structural texture similarity metrics validated the performance of the proposed subjective testing procedure and the metrics evaluated.

Conclusions:

• ViSiProG offers an efficient and effective method for conducting subjective experiments to validate texture similarity metrics.
• The defined performance requirements and testing domains provide a framework for developing and evaluating objective texture similarity metrics.
• The study highlights the importance of aligning objective metrics with human perceptual judgments for robust image analysis applications.