Multi-Interactive Dual-Decoder for RGB-Thermal Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• RGB-thermal salient object detection (RGBT SOD) aims to identify common prominent objects in both visible and thermal infrared images.
• Current RGBT SOD methods often fail to fully exploit the complementary nature of different modalities and diverse image cues.
• This limitation hinders the achievement of highly accurate salient object segmentation.

Purpose of the Study:

• To propose a novel multi-interactive dual-decoder architecture for enhanced RGBT SOD.
• To effectively mine and model multi-type interactions between RGB and thermal modalities.
• To improve the accuracy and robustness of salient object detection in challenging conditions.

Main Methods:

• Encoding RGB and thermal images into multi-level, multi-modal feature representations.
• Designing a novel dual-decoder to facilitate interactions among multi-level features, modalities, and global contexts.
• Implementing extensive experiments on public RGBT and RGBD SOD datasets.

Main Results:

• The proposed multi-interactive dual-decoder demonstrates superior performance in RGBT SOD.
• The method achieves outstanding results compared to existing state-of-the-art algorithms.
• Effective handling of challenging scenarios, including cases with invalid modalities.

Conclusions:

• The proposed multi-interactive dual-decoder effectively addresses limitations in current RGBT SOD approaches.
• The method's ability to model multi-modal interactions leads to significant performance gains.
• The approach shows strong potential for real-world applications requiring accurate salient object detection.