Bio-Inspired Representation Learning for Visual Attention Prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Visual attention prediction (VAP) is crucial in computer vision.
• Existing deep learning VAP methods often overlook low-level contrast features.
• Human visual perception relies on both contrast and semantic information.

Purpose of the Study:

• To propose a novel VAP method leveraging bio-inspired representation learning.
• To integrate low-level contrast and high-level semantic features for improved attention mapping.
• To enhance the accuracy of visual attention maps in computer vision applications.

Main Methods:

• Feature extraction using refined VGG16 for semantic features and a novel contrast feature extraction block for low-level features.
• Bio-inspired representation learning combining contrast and semantic features via a densely connected block.
• Visual attention map generation using a weighted-fusion layer.

Main Results:

• The proposed method effectively integrates low-level contrast and high-level semantic features.
• Experimental results demonstrate the superiority of the novel VAP approach.
• The bio-inspired representation learning enhances the generation of visual attention maps.

Conclusions:

• The novel bio-inspired VAP method significantly improves attention map generation.
• Combining contrast and semantic features is key to mimicking human visual attention.
• This approach offers a promising direction for future VAP research.