PSRNet: a phase-guided frequency-domain structure reconstruction network for RGB-T salient object detection

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Salient object detection in RGB-Thermal (RGB-T) data is difficult due to structural shifts and weak thermal boundaries.
• Existing methods struggle with cross-modal feature alignment and precise boundary delineation.

Purpose of the Study:

• To propose PSRNet, a phase-guided frequency-domain structure reconstruction network for robust RGB-T salient object detection.
• To address challenges of structural shifts, weak thermal boundaries, and background interference in RGB-T data.

Main Methods:

• Decomposes RGB and thermal features into amplitude and latent-phase components in the Fourier domain.
• Aligns cross-modal structural cues using amplitude-weighted phase consistency.
• Reconstructs high-frequency boundary information with a bounded Gaussian high-pass gate, followed by adaptive phase-modulated fusion and multi-scale supervision.

Main Results:

• Achieved an S-measure of 0.903, MAE of 0.028, and F-measure of 0.794 (at threshold 0.80) on the VT1000 dataset.
• Obtained a boundary IoU of 0.862, demonstrating superior boundary recovery.
• Validated improved structural preservation and boundary delineation under challenging RGB-T conditions.

Conclusions:

• PSRNet effectively overcomes limitations in RGB-T salient object detection.
• The proposed frequency-domain approach significantly enhances structural and boundary accuracy.
• Results indicate a promising direction for future RGB-T perception tasks.