GDVIFNet: A generated depth and visible image fusion network with edge feature guidance for salient object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Salient Object Detection (SOD) performance is suboptimal in complex environments.
• Existing SOD methods often require additional depth or thermal sensors, increasing cost and inconvenience.

Purpose of the Study:

• To propose GDVIFNet, a novel approach for salient object detection using only a single RGB image.
• To overcome the limitations of specialized hardware by leveraging generated depth information.

Main Methods:

• Utilizes Depth Anything to generate depth images from RGB input.
• Incorporates self-supervised techniques to generate edge features for noise and artifact removal from generated depth images.
• Employs a dual-branch architecture (CNN and Transformer) with novel fusion modules (STIU, SCF, CETSF) for feature integration.

Main Results:

• GDVIFNet achieves state-of-the-art performance on multiple salient object detection datasets.
• The proposed method effectively handles noise and artifacts in generated depth data.
• Demonstrates the efficacy of fusing RGB, generated depth, and edge features.

Conclusions:

• GDVIFNet offers a cost-effective and convenient solution for salient object detection.
• The integration of generated depth and self-supervised edge features significantly enhances SOD performance.
• The dual-branch architecture with advanced fusion mechanisms proves effective for complex visual scenes.