Salient Region Detection Using Diffusion Process on a Two-Layer Sparse Graph.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Diffusion-based salient region detection is a key area in computer vision.
• Existing methods face challenges with diffusion matrix construction and seed vector generation, especially for objects near image boundaries.

Purpose of the Study:

• To improve the accuracy and robustness of diffusion-based salient region detection.
• To introduce novel approaches for diffusion matrix and seed vector generation.

Main Methods:

• Constructed a two-layer sparse graph connecting nodes based on local neighborhoods and boundary similarity.
• Utilized spatial variance of superpixel clusters for seed vector generation to better distinguish saliency from background.
• Integrated two preliminary saliency maps derived from saliency and background seeds using manifold ranking diffusion.

Main Results:

• The proposed graph construction effectively captures local spatial relationships while removing redundant nodes.
• The new seed vector approach improves saliency seed distinction, particularly for boundary-located objects.
• Extensive experiments demonstrated superior performance compared to 20 state-of-the-art methods on five benchmark datasets.

Conclusions:

• The enhanced diffusion-based method significantly improves salient region detection accuracy.
• The novel graph construction and seed vector generation offer a more robust approach to saliency detection.
• The method shows strong potential for various image analysis applications.