Scene Categorization Using Deeply Learned Gaze Shifting Kernel.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cognitive Science

Background:

• Scene recognition is crucial for intelligent systems like autonomous driving.
• Current deep learning models lack explicit encoding of human visual perception and gaze patterns.
• Understanding how humans perceive scenes is key to improving recognition accuracy.

Purpose of the Study:

• To develop a novel deep gaze shifting kernel for enhanced scene categorization.
• To explicitly model human visual perception by analyzing gaze allocation sequences.
• To improve the accuracy and robustness of scene recognition systems.

Main Methods:

• Projecting scene regions into a perceptual space using combined features (color, texture, semantic).
• Employing a novel non-negative matrix factorization algorithm to derive region saliency and gaze paths.
• Utilizing an aggregation-based convolutional neural network to learn deep representations of gaze paths.
• Integrating deep representations into an image kernel for kernel Support Vector Machine (SVM) classification.

Main Results:

• The proposed deep gaze shifting kernel method significantly outperforms existing shallow and deep recognition models across six datasets.
• Predicted gaze shifting paths demonstrate high consistency (94.6%) with actual human eye-tracking data.
• The approach effectively captures human visual perception patterns for scene understanding.

Conclusions:

• The deep gaze shifting kernel offers a superior approach to scene categorization by incorporating human visual perception.
• Modeling gaze allocation paths provides a more nuanced understanding of scene complexity.
• This method represents a significant advancement in intelligent systems requiring accurate scene recognition.