SG-FCN: A Motion and Memory-Based Deep Learning Model for Video Saliency Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Convolutional Neural Networks (CNNs) have advanced image saliency detection.
• Video fixation detection remains underexplored, lacking models that integrate motion and temporal dynamics.
• Existing methods struggle with complex scenes due to insufficient inter-frame analysis.

Purpose of the Study:

• To develop a novel and efficient video eye fixation detection model.
• To enhance saliency detection performance in video sequences.
• To address limitations of current models in handling temporal and motion information.

Main Methods:

• A step-gained fully convolutional network (FCN) was proposed.
• The model simulates human memory and visual attention mechanisms.
• It combines temporal memory information with spatial motion information, storing current frame saliency.

Main Results:

• The proposed model demonstrated superior performance compared to 11 state-of-the-art methods.
• Experiments were conducted on multiple public datasets.
• Hierarchical training ensured high detection accuracy.

Conclusions:

• The novel model significantly improves video eye fixation detection.
• Integrating motion and temporal information is key for accurate video saliency.
• The approach offers a promising direction for future research in video attention modeling.