Curiosity-Driven Salient Object Detection With Fragment Attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods with attention mechanisms excel in salient object detection.
• Existing attention mechanisms suffer from computational redundancy or insufficient targeting, leading to inaccuracies.

Purpose of the Study:

• To address limitations in current attention mechanisms for salient object detection.
• To introduce a novel Curiosity-driven Network (CNet) and algorithm (CLA) for improved performance.

Main Methods:

• Developed a fragment attention (FA) mechanism inspired by human curiosity, with three curiosity levels.
• Proposed a high-level feature extraction module (HFEM) for context-aware information.
• Implemented a Curiosity-driven Learning Algorithm (CLA) to transform pixel curiosity into detailed saliency maps.

Main Results:

• The proposed FA mechanism accurately classifies pixel curiosity based on enhanced high-level features from HFEM.
• Extensive experiments on five datasets show superior performance compared to state-of-the-art methods.
• The method achieves high accuracy without pre- or post-processing operations.

Conclusions:

• The CNet and CLA, powered by the FA mechanism and HFEM, significantly advance salient object detection.
• This novel approach offers a more accurate and efficient alternative to existing attention-based methods.
• The curiosity-driven paradigm provides a promising direction for future research in computer vision.