Perceptually Motivated Image Features Using Contours

Area of Science:

• Computer Vision
• Human Perception
• Image Analysis

Background:

• Existing computational feature sets struggle to accurately estimate human perceptual texture similarity.
• Current methods often overlook the human visual system's ability to utilize long-range contour information.
• Higher-order statistics (HOS) over larger image regions are underexplored for texture analysis.

Purpose of the Study:

• To investigate the utility of long-range HOS, specifically contour data, for perceptual texture similarity estimation.
• To develop a novel set of perceptually motivated image features (PMIF) that leverage these long-range HOS.
• To compare the performance of PMIF against existing feature sets and contour representations.

Main Methods:

• Conducted a psychophysical experiment to assess the importance of contour data versus local image patches and global second-order data.
• Developed PMIF by encoding long-range HOS from spatial and angular distributions of contour segments.
• Evaluated PMIF on perceptual texture similarity, sketch-based image retrieval, and natural scene recognition tasks.

Main Results:

• Psychophysical data indicate contour data are more critical for human texture perception than local or global features.
• PMIF demonstrated superior or comparable performance against 51 existing feature sets and four contour representations.
• The proposed features showed effectiveness in multiple image analysis tasks beyond texture similarity.

Conclusions:

• Long-range HOS, particularly contour data, are crucial for accurate perceptual texture similarity estimation.
• The proposed perceptually motivated image features (PMIF) effectively capture both short-range and long-range HOS.
• PMIF offers a promising new approach for various image analysis applications, including retrieval and recognition.