Advanced Discriminative Co-Saliency and Background Mining Transformer for Co-Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing co-salient object detection (CoSOD) models often overlook background regions, hindering performance in complex environments.
• This limitation can lead to difficulties in accurately identifying salient objects when background interference is significant.

Purpose of the Study:

• To propose a novel Discriminative co-saliency and background Mining Transformer (DMT) framework.
• To explicitly mine both co-saliency and background information for improved discriminability.
• To enhance the robustness and practicality of CoSOD models, particularly in open-world scenarios.

Main Methods:

• Developed DMT framework utilizing disjoint extraction of co-saliency and background tokens from segmentation features.
• Introduced economic multi-grained correlation modules (R2R, CtP2T, CoT2T) for efficient information extraction.
• Implemented Token-Guided Feature Refinement (TGFR) modules, enhanced to Group TGFR (G-TGFR), and a Noise Propagation Suppression (NPS) mechanism for DMT+O.

Main Results:

• The proposed DMT framework effectively mines both co-saliency and background information, improving discriminability.
• Experimental results demonstrate superior performance on conventional and open-world CoSOD benchmark datasets.
• The extended DMT+O version shows enhanced practicality and effectiveness in real-world applications.

Conclusions:

• The DMT framework offers a significant advancement in CoSOD by addressing the limitations of neglecting background information.
• The proposed methods, including G-TGFR and NPS, enhance model discriminability and applicability.
• DMT provides a more robust and effective solution for co-salient object detection tasks.