SLMSF-Net: A Semantic Localization and Multi-Scale Fusion Network for RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Salient Object Detection (SOD) in RGB-D images is vital for computer vision.
• Optimizing multi-modal and multi-scale feature fusion for enhanced SOD performance is challenging.

Purpose of the Study:

• To propose a novel network model, SLMSF-Net, for RGB-D SOD.
• To improve the fusion of multi-modal and multi-scale features for precise salient object identification.

Main Methods:

• Developed a Deep Attention Module (DAM) for merging RGB and depth features.
• Introduced a Semantic Localization Module (SLM) for precise object localization.
• Employed a Multi-Scale Fusion Module (MSF) for detailed information restoration and high-precision saliency map generation.

Main Results:

• SLMSF-Net demonstrated improved performance across six RGB-D datasets.
• Achieved performance gains of 0.20-1.80% in maxF, 0.09-1.46% in maxE, 0.19-1.05% in S, and 0.0002-0.0062 in MAE.
• Outperformed competing methods like AFNet, DCMF, and C2DFNet.

Conclusions:

• SLMSF-Net effectively addresses the challenges in RGB-D salient object detection.
• The proposed modules enhance feature fusion, localization, and detail restoration for superior saliency map generation.