DMGNet: Depth mask guiding network for RGB-D salient object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Depth images offer spatial cues for salient object detection (SOD).
• Improper use of depth features can degrade SOD performance by introducing noise.
• Existing methods struggle to effectively integrate RGB and depth data for SOD.

Purpose of the Study:

• To propose a novel network, the Depth Mask Guiding Network (DMGNet), for robust RGB-D salient object detection.
• To enhance the extraction of discriminative features by guiding the RGB subnetwork with depth-derived masks.
• To improve the fusion of cross-modal features for more accurate salient object detection.

Main Methods:

• A Depth Mask Guidance Module (DMGM) pre-segments salient objects from depth images to generate guiding masks.
• A Feature Fusion Pyramid Module (FFPM) with multi-branch convolutional channels fuses cross-modal features.
• The proposed DMGNet integrates depth-based guidance and enhanced feature fusion for RGB-D SOD.

Main Results:

• The DMGNet effectively leverages depth information, overcoming limitations of noisy or misleading depth features.
• The DMGM successfully guides the RGB subnetwork, leading to more discriminative feature extraction.
• The FFPM enhances the fusion of RGB and depth features, improving overall SOD accuracy.
• Extensive experiments on nine benchmark datasets validate the proposed network's effectiveness.

Conclusions:

• The proposed DMGNet significantly improves salient object detection performance in RGB-D scenarios.
• Effective utilization of depth information through mask guidance is crucial for robust SOD.
• The network architecture offers a promising approach for advanced RGB-D image analysis tasks.