DMRA: Depth-Induced Multi-Scale Recurrent Attention Network for RGB-D Saliency Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• RGB-D saliency detection aims to identify visually important regions in images using both color (RGB) and depth information.
• Complex scenes pose challenges due to intricate details and overlapping objects, requiring advanced feature extraction and fusion techniques.

Purpose of the Study:

• To propose a novel RGB-D saliency detection network, DMRA, that improves accuracy, particularly in complex scenarios.
• To introduce a new benchmark RGB-D saliency dataset for evaluating model performance.

Main Methods:

• A depth refinement block with residual connections for cross-modal feature fusion.
• Integration of depth cues with multi-scale contextual features for precise object localization.
• A recurrent attention module inspired by the brain's internal generative mechanism for semantic understanding and detail optimization.
• A cascaded hierarchical feature fusion strategy for efficient multi-level feature interaction.

Main Results:

• DMRA demonstrates significant performance improvements, especially in complex visual environments.
• The proposed method accurately identifies salient objects.
• Experimental results show superior performance compared to 18 state-of-the-art RGB-D saliency models across nine benchmark datasets.

Conclusions:

• The DMRA network offers a robust solution for RGB-D saliency detection, effectively handling complex scenes.
• The novel architectural components and fusion strategies contribute to enhanced accuracy and contextual representability.
• The introduced dataset serves as a valuable resource for advancing RGB-D saliency research.