An improved saliency model of visual attention dependent on image content

Area of Science:

• Computer Vision
• Cognitive Science
• Computational Neuroscience

Background:

• Existing visual attention models often overlook crucial mechanisms like feature informativeness, scale integration, and image structure's influence on attention.
• These omissions lead to models that do not accurately fit human visual data.

Purpose of the Study:

• To develop a more accurate visual attention model by integrating underrepresented mechanisms.
• To improve the prediction of image saliency maps that align with human perception.

Main Methods:

• Incorporated medium-level features (orientation and color combinations) alongside low-level features.
• Utilized a variable number of center-surround difference maps to account for image structure diversity.
• Integrated information from different scales and features using a weighted sum approach, considering local and global saliency.

Main Results:

• The proposed model demonstrated high performance on the CAT2000 dataset, achieving an Area Under Curve (AUC) of 0.79 and a saliency AUC (sAUC) of 0.58.
• The model's results indicate a better fit to human visual attention data compared to previous models.

Conclusions:

• The integrated mechanisms, including medium-level features and adaptive scale processing, significantly enhance visual saliency prediction.
• The proposed model offers a framework for improving existing visual attention models by incorporating these biologically plausible mechanisms.