Visual Saliency via Embedding Hierarchical Knowledge in a Deep Neural Network.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) are widely used in image processing, including visual saliency prediction.
• A key challenge in DNN-based saliency prediction is the scarcity of labeled ground truth data, leading to model overfitting due to numerous trainable parameters.

Purpose of the Study:

• To develop a novel method for visual saliency prediction that overcomes the limitations of limited ground truth data.
• To reduce model overfitting in deep neural networks for saliency prediction by leveraging existing knowledge.

Main Methods:

• A novel DNN method embedding hierarchical knowledge from existing visual saliency models (local, global, semantic).
• Tuning and fixing approximately 92.5% of network parameters hierarchically using pre-generated saliency maps.
• Incorporating a center prior into the learning cost function, emphasizing errors near the image center.

Main Results:

• Significantly reduced the number of trainable parameters requiring ground truth tuning.
• Enabled full utilization of large DNN capabilities while mitigating overfitting.
• Demonstrated superior performance compared to classical and state-of-the-art methods across various evaluation metrics on four public databases.

Conclusions:

• The proposed method effectively addresses the challenge of limited ground truth data in visual saliency prediction.
• Embedding hierarchical knowledge and incorporating a center prior are effective strategies for improving DNN performance and robustness.
• The method offers a significant advancement in accurate and reliable visual saliency map prediction.