Fast and robust generation of feature maps for region-based visual attention

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computational Neuroscience

Background:

• Visual attention is crucial for efficient biological and artificial vision.
• Current artificial attention models often employ late clustering, limiting feature reusability.
• There is a need for region-based attention mechanisms that preserve spatial information.

Purpose of the Study:

• To propose a region-based visual attention model with pre-attentive pixel clustering.
• To introduce novel algorithms for constructing color contrast, symmetry, and size contrast feature maps.
• To evaluate the model's efficiency and the reusability of salient regions in machine vision.

Main Methods:

• Region-based approach with early pixel clustering.
• Construction of color contrast map using color theory.
• Novel scanning-based method for symmetry map generation.
• New algorithm for size contrast map computation.
• Saliency evaluation based on rarity criteria using region moments.

Main Results:

• The model efficiently incorporates five feature channels.
• Processing rates of multiple frames per second are maintained.
• Preservation of salient region shapes and locations enables reusability.
• Demonstrated potential for integrating attention into mainstream machine vision.

Conclusions:

• The proposed region-based attention model offers an efficient alternative to late clustering methods.
• The model's ability to preserve spatial information enhances feature reusability for downstream tasks.
• Systems with limited computational resources, like mobile robots, can significantly benefit from this approach.