Salient object detection in low-light RGB-T scene via spatial-frequency cues mining

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light conditions severely impair vision tasks like salient object detection (SOD) due to limited photons.
• Existing RGB-Thermal SOD models offer partial solutions but neglect frequency discrepancies, limiting their performance.

Purpose of the Study:

• To develop an advanced RGB-Thermal SOD model, SFMNet, specifically for low-light scenes.
• To address the limitations of current models by incorporating spatial-frequency cues.

Main Methods:

• Proposed SFMNet model integrating spatial-frequency feature exploration (SFFE) and spatial-frequency feature interaction (SFFI) modules.
• SFFE modules adaptively extract high and low-frequency features, while SFFI modules fuse cross-modality and cross-domain information.
• A top-down pathway architecture was employed for high-quality saliency prediction.

Main Results:

• SFMNet demonstrated superior performance in low-light RGB-T SOD tasks compared to existing models.
• The model achieved higher accuracy, indicating the effectiveness of spatial-frequency cue mining.
• The first low-light RGB-T SOD dataset was created as a benchmark.

Conclusions:

• SFMNet offers a significant advancement in low-light salient object detection.
• The integration of spatial-frequency cues is crucial for enhancing SOD performance in challenging lighting.
• The developed dataset will facilitate future research in low-light computer vision.