Position-Aware Relation Learning for RGB-Thermal Salient Object Detection.

Area of Science:

• Computer Vision
• Image Segmentation
• Machine Learning

Background:

• Salient Object Detection (SOD) identifies conspicuous image regions.
• RGB-Thermal SOD leverages both visual and thermal spectra for enhanced segmentation.
• Existing methods struggle with boundary details due to isolated pixel considerations.

Purpose of the Study:

• To propose a novel Position-Aware Relation Learning Network (PRLNet) for RGB-Thermal SOD.
• To address limitations in current methods that ignore pixel interactions for boundary refinement.
• To improve the accuracy and robustness of salient object detection in RGB-Thermal imagery.

Main Methods:

• Developed a Position-Aware Relation Learning Network (PRLNet) incorporating distance and direction relationships.
• Introduced a Signed Distance Map Auxiliary Module (SDMAM) to enhance feature representation and inter-class separation.
• Implemented a Feature Refinement approach with Direction Field (FRDF) to improve intra-class compactness of salient features.
• Utilized a transformer-based decoder for effective multispectral feature fusion.

Main Results:

• PRLNet significantly outperforms state-of-the-art methods on three public RGB-T SOD datasets.
• The method demonstrates improved intra-class compactness and inter-class separation.
• Ablation studies and visualizations validate the effectiveness and interpretability of the proposed approach.

Conclusions:

• The proposed PRLNet effectively addresses limitations in existing RGB-T SOD methods.
• The network achieves superior performance by modeling pixel relationships through distance and direction.
• PRLNet offers a flexible, plug-and-play solution compatible with various backbone networks.