A global/local affinity graph for image segmentation.

Area of Science:

• Computer Vision
• Image Processing
• Computational Neuroscience

Background:

• Graph-cut based image segmentation relies on accurate graphs representing perceptual grouping cues.
• Existing methods often struggle to capture both short- and long-range dependencies effectively.

Purpose of the Study:

• To propose a novel sparse global/local affinity graph for improved image segmentation.
• To capture diverse perceptual grouping laws (proximity, similarity, continuity) using superpixels.
• To evaluate and fuse color, texture, and shape features for enhanced perceptual grouping.

Main Methods:

• Image oversegmentation into superpixels at multiple scales.
• Adaptive superpixel division into small, medium, and large sets based on a gravitation law.
• Sparse representation using l0-minimization for global grouping (continuity) with medium superpixels.
• Adjacent graph construction for local smoothness (similarity, proximity) with small and large superpixels.
• Introduction of a bipartite graph for inter-scale superpixel cue propagation.

Main Results:

• The proposed sparse graph effectively captures both short- and long-range grouping cues.
• Fusion of color, texture, and shape features, guided by psychophysics, improves segmentation.
• Experimental results on the Berkeley segmentation database demonstrate superior performance.
• Outperforms state-of-the-art graph constructions across four objective criteria (PRI, VI, GCE, BDE).

Conclusions:

• The novel sparse global/local affinity graph significantly advances graph-cut based image segmentation.
• The method's ability to integrate multi-scale superpixels and diverse visual features leads to robust perceptual grouping.
• This approach offers a promising direction for more accurate and perceptually relevant image segmentation.