Potential region attention network for RGB-D salient object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Existing RGB-D salient object detection (SOD) methods often struggle with effectively utilizing complementary features from different modalities.
• Limited mining of single-modal features and suboptimal cross-modal feature fusion hinder performance in current approaches.

Purpose of the Study:

• To propose a novel Potential Region Attention Network (PRANet) for enhanced RGB-D salient object detection (SOD).
• To improve feature representation by effectively leveraging both intra-modal and cross-modal information.

Main Methods:

• Employs Swin Transformer backbone for efficient two-stream feature extraction.
• Introduces a Potential Multi-scale Attention Module (PMAM) for enhanced intra-modal feature expression.
• Designs a Potential Region Attention Module (PRAM) to guide effective cross-modal feature fusion.
• Incorporates a Feature Refinement Fusion Module (FRFM) to strengthen encoder-decoder information transmission.
• Utilizes multi-side supervision during the training phase.

Main Results:

• PRANet demonstrates superior performance compared to 15 representative methods.
• Achieves outstanding results across six diverse RGB-D SOD datasets.
• Validated effectiveness of PMAM and PRAM in enhancing feature mining and fusion.

Conclusions:

• PRANet effectively addresses limitations in single-modal feature mining and cross-modal feature complementarity.
• The proposed attention mechanisms significantly improve RGB-D salient object detection performance.
• PRANet represents a significant advancement in the field of RGB-D SOD.