Decomposition and Completion Network for Salient Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Fully convolutional networks (FCNs) have advanced salient object detection.
• Current methods often focus on integrating edge information into deep models.
• There is a need for improved methods that model object integrity.

Purpose of the Study:

• To propose a novel Decomposition and Completion Network (DCN) for salient object detection.
• To integrate edge and skeleton information as complementary features.
• To model the integrity of salient objects in a two-stage process.

Main Methods:

• Developed a cross multi-branch decoder for integrating multi-level, multi-task features.
• Predicted saliency, edge, and skeleton maps simultaneously in the decomposition network.
• Utilized edge and skeleton maps for noise suppression and flaw filling in the completion network via hierarchical structure-aware learning and multi-scale feature completion.

Main Results:

• The proposed DCN generates precise saliency maps with uniformly and completely segmented salient objects.
• Experiments on five benchmark datasets show superior performance compared to existing networks.
• The model extended to RGB-D salient object detection also achieved state-of-the-art results.

Conclusions:

• The DCN effectively leverages complementary edge and skeleton information for accurate salient object detection.
• Jointly learning boundary and interior information leads to improved saliency map generation.
• The proposed approach represents a significant advancement in salient object detection and its extension to RGB-D data.