MSEDNet: Multi-scale fusion and edge-supervised network for RGB-T salient object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Salient Object Detection (SOD) aims to segment important image regions.
• Existing SOD methods often overlook the complementary information between visible light and thermal infrared images.
• Improved SOD accuracy can benefit various applications, including surveillance and autonomous systems.

Purpose of the Study:

• To introduce MSEDNet, a novel method for RGB-Thermal (RGB-T) Salient Object Detection (SOD).
• To effectively leverage the complementary information between visible light and thermal infrared images for enhanced SOD.
• To improve the accuracy and robustness of salient object detection in multi-modal scenarios.

Main Methods:

• Utilized an encoder to extract multi-level features (high, medium, low) from both RGB and Thermal images.
• Proposed three specialized feature fusion modules: Edge Dilation Sharpening (low-level), Spatial and Channel-Aware (mid-level), and Cross-Residual Fusion (high-level).
• Introduced an edge fusion loss function for supervised learning, focusing on edge information extraction and noise suppression.

Main Results:

• MSEDNet demonstrated superior performance compared to existing state-of-the-art SOD methods in comparative evaluations.
• The proposed multi-level feature fusion strategy effectively captured complementary information between RGB and Thermal data.
• The edge fusion loss function contributed to precise boundary delineation and background noise reduction.

Conclusions:

• MSEDNet offers a significant advancement in RGB-T salient object detection by effectively fusing multi-modal information.
• The method's architecture and loss function provide a robust framework for accurate salient object segmentation.
• The findings highlight the importance of cross-modal feature complementarity for improving SOD performance.