A perceptually based comparison of image similarity metrics

Area of Science:

• Computer Vision
• Human Visual Perception
• Image Analysis

Background:

• Image similarity assessment is crucial for human visual processing models and image analysis systems.
• L1 and L2 norms, instances of the Minkowski metric, are commonly used but lack a clear principled reason for selection.
• Understanding which metric better aligns with human perception can inform both neuroscience and computer vision applications.

Purpose of the Study:

• To investigate whether the L1 or L2 norm better captures the human perceptual notion of image similarity.
• To derive inferences about human visual system similarity criteria.
• To evaluate and potentially improve image analysis metrics.

Main Methods:

• Participants evaluated perceptual preferences for images retrieved using L1 versus L2 norms.
• Images used were either small, content-less fragments or larger, recognizable patterns generated by vector quantization.
• A comparative analysis of user preferences between L1 and L2 metric-based image retrieval was conducted.

Main Results:

• Participants exhibited a small but consistent preference for images retrieved using the L1 metric over the L2 metric.
• This preference was observed across both content-less image fragments and recognizable vector-quantized patterns.
• The findings indicate a potential advantage of the L1 norm in reflecting human similarity judgments for natural images.

Conclusions:

• The L1 metric may provide a more accurate representation of human notions of image similarity for the tested natural image domains.
• These findings have implications for developing more perceptually aligned image analysis algorithms.
• Further research could explore the generalizability of these findings across diverse image types and contexts.