Swin Transformer-Based Edge Guidance Network for RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Salient Object Detection (SOD) is crucial for computer vision.
• Existing RGB-D SOD methods using Convolutional Neural Networks (CNNs) suffer from limited performance due to inherent locality.
• There is a need for methods that can capture global context and effectively fuse multi-modal features.

Purpose of the Study:

• To propose a novel Swin Transformer-based edge guidance network (SwinEGNet) for RGB-D SOD.
• To address the limitations of CNN-based methods by incorporating global context extraction.
• To enhance feature fusion through an edge-guided cross-modal interaction module.

Main Methods:

• Employed Swin Transformer as a backbone for feature extraction from RGB and depth images.
• Introduced an Edge Extraction Module (EEM) and a Depth Enhancement Module (DEM).
• Utilized a Cross-Modal Interaction Module (CIM) for integrating global and local cross-modal features, followed by a cascaded decoder for refinement.

Main Results:

• SwinEGNet achieved state-of-the-art performance on LFSD, NLPR, DES, and NJU2K datasets.
• Demonstrated comparable performance on the STEREO dataset against 14 other methods.
• Achieved superior performance over SwinNet with significantly fewer parameters (88.4%) and FLOPs (77.2%).

Conclusions:

• SwinEGNet effectively captures global context and enhances feature fusion for RGB-D SOD.
• The proposed edge-guided approach significantly improves detection accuracy and efficiency.
• The model's performance and efficiency suggest its potential for real-world computer vision applications.