Quality-Aware Selective Fusion Network for V-D-T Salient Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Salient Object Detection (SOD) benefits from multi-modal data like RGB, depth, and thermal images.
• Existing methods struggle with unreliable depth and thermal image quality, degrading SOD performance.
• Triple-modal SOD (VDT) methods often overlook the quality of input modalities.

Purpose of the Study:

• To propose a Quality-Aware Selective Fusion Network (QSF-Net) for VDT salient object detection.
• To address performance degradation caused by low-quality depth and thermal images.
• To enhance the fusion of multi-modal features by considering image quality.

Main Methods:

• Developed a QSF-Net with three subnets: initial feature extraction, quality-aware region selection, and region-guided selective fusion.
• Employed a weakly-supervised approach to generate quality-aware maps using preliminary predictions.
• Integrated multi-scale fusion, intra- and inter-modality attention, and edge refinement modules.

Main Results:

• The proposed QSF-Net demonstrates superior performance in VDT salient object detection.
• The model consistently outperformed 13 state-of-the-art methods on the VDT-2048 dataset.
• Effective handling of low-quality depth and thermal images was achieved.

Conclusions:

• QSF-Net offers a robust solution for VDT salient object detection, particularly in challenging conditions.
• Quality-aware fusion is crucial for improving the reliability of multi-modal SOD.
• The proposed method advances the state-of-the-art in triple-modal salient object detection.