Combining segmentation and attention: a new foveal attention model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computational Neuroscience

Background:

• Artificial vision systems face computational challenges in real-time processing.
• Biological perception, particularly human foveal attention, offers a model for efficient visual processing.
• Attention and segmentation are bidirectionally linked in human vision, with attention guiding segmentation.

Purpose of the Study:

• To develop a bottom-up foveal attention model for artificial vision systems.
• To integrate multi-resolution perceptual segmentation with saliency estimation.
• To improve the efficiency and accuracy of artificial visual information processing.

Main Methods:

• Input images are represented using Cartesian Foveal Geometry (CFG) for multi-resolution perception.
• Perceptual segmentation is achieved by constructing a foveal polygon using the Bounded Irregular Pyramid (BIP).
• Bottom-up attention is integrated to identify salient proto-objects based on features like color, intensity, symmetry, orientation, and roundness.

Main Results:

• The proposed model successfully integrates hierarchical foveal segmentation and saliency estimation.
• Experimental results on natural images demonstrate good performance in terms of accuracy.
• The model achieves significant improvements in processing speed compared to traditional methods.

Conclusions:

• The developed bottom-up foveal attention model effectively mimics biological visual perception.
• This approach offers a computationally efficient solution for artificial vision systems.
• The integration of foveal segmentation and saliency estimation enhances real-time visual processing capabilities.