Saliency Detection via Absorbing Markov Chain With Learnt Transition Probability.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Traditional saliency detection models often struggle with uniform highlighting of salient objects.
• Accurate representation of local context and global object importance is crucial for effective saliency mapping.

Purpose of the Study:

• To propose a novel bottom-up saliency model leveraging absorbing Markov chains (AMC) for improved image saliency detection.
• To enhance the uniform highlighting of salient objects by refining saliency maps.

Main Methods:

• Constructing a sparsely connected graph and employing an absorbing Markov chain (AMC) model where boundary nodes are absorbing and others are transient.
• Calculating saliency values based on the expected number of transitions between transient nodes, encoded in a learned transition probability matrix derived from deep features.
• Utilizing an angular embedding technique to refine saliency results by rearranging global orderings based on pairwise local orderings.

Main Results:

• The proposed AMC-based model significantly enhances performance compared to state-of-the-art methods.
• The learned transition probability matrix, incorporating deep features, improves the saliency detection accuracy.
• Angular embedding effectively refines saliency maps for more uniform highlighting of salient objects.

Conclusions:

• The developed absorbing Markov chain model offers a robust approach to bottom-up saliency detection.
• The integration of deep features and angular embedding provides a significant advancement in highlighting salient objects uniformly.
• The model demonstrates superior performance across multiple benchmark datasets, validating its effectiveness.