Scene Categorization by Deeply Learning Gaze Behavior in a Semisupervised Context.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cognitive Science

Background:

• Deep learning models achieve high accuracy in scene categorization but lack human visual perception characterization.
• Understanding human gaze allocation and cognitive processes is crucial for advanced scene understanding.

Purpose of the Study:

• To propose a novel spatially aware aggregation network for scene categorization.
• To discover human gaze behavior in a semisupervised setting for improved scene recognition.

Main Methods:

• Developed a semisupervised, structure-preserved non-negative matrix factorization (NMF) to identify salient regions.
• Engineered a gaze shifting path (GSP) to model human visual perception sequences.
• Created a spatially aware CNN (SA-Net) to analyze GSP features and aggregated salient regions.

Main Results:

• The proposed SA-Net effectively characterizes gaze shifting paths for scene perception.
• Fusion of GSP features with kernel SVM demonstrated superior scene categorization performance.
• Comparative experiments on six datasets validated the method's advantage over existing approaches.

Conclusions:

• The developed spatially aware aggregation network offers a novel approach to scene categorization by incorporating human visual perception.
• The semisupervised discovery of gaze behavior and SA-Net's deep feature extraction significantly enhance scene understanding capabilities.
• This method provides a more cognitively plausible and accurate framework for computer vision tasks like autonomous driving.