Cross-modal interactive and global awareness fusion network for RGB-D salient object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• RGB-D salient object detection leverages depth data for enhanced performance over RGB-only methods.
• Existing models struggle with complex scenes, multiple objects, and small object detection.

Purpose of the Study:

• To propose a novel Cross-modal Interactive and Global Awareness Fusion Network (CIGNet) for robust RGB-D salient object detection.
• To enhance the accurate recognition and highlighting of salient objects in complex visual scenes.

Main Methods:

• Developed CIGNet integrating Convolutional Neural Networks (CNNs) and attention mechanisms for RGB and depth data fusion.
• Introduced Cross-modal Interaction Fusion Module (CIFM) using depth separable and dynamic convolutions for detailed feature extraction.
• Designed Global Awareness Fusion Module (GAFM) to integrate high-level RGB and depth features for improved scene understanding.
• Utilized Multi-layer Convolutional Fusion Module (MCFM) for a step-by-step decoding process to refine detection results.

Main Results:

• CIGNet demonstrates superior robustness and accuracy compared to 12 mainstream methods.
• The proposed fusion methods effectively extract local details and integrate global information.
• The model shows improved performance in complex scenes with multiple or small salient objects.

Conclusions:

• CIGNet significantly advances RGB-D salient object detection capabilities.
• The network's architecture effectively addresses limitations of existing models in challenging scenarios.
• The findings suggest CIGNet is a highly effective solution for accurate salient object detection.