Capturing the grouping and compactness of high-level semantic feature for saliency detection

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Saliency detection is challenging due to complex backgrounds and regions.
• Existing methods often require supervision or struggle with feature complexity.

Purpose of the Study:

• To develop a novel unsupervised saliency detection approach.
• To leverage high-level semantic features for improved accuracy and efficiency.

Main Methods:

• Utilizing an elastic net-based hypergraph model to identify group structures in salient regions.
• Employing spatial distribution calculations to determine the compactness of saliency regions.
• Implementing a propagation algorithm with an enhanced similarity matrix fusing low-level deep and high-level semantic features.

Main Results:

• Demonstrated effectiveness on four benchmark datasets with pixel-wise accurate labeling.
• Achieved competitive performance against supervised deep learning-based methods.
• Validated the efficacy of exploiting grouping and compactness characteristics.

Conclusions:

• The proposed unsupervised method offers a viable alternative to supervised approaches.
• High-level semantic features are crucial for robust saliency detection.
• The fusion of low-level and high-level features enhances saliency map quality.