Bio-inspired two-stage network for efficient RGB-D salient object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) and Vision Transformers have advanced RGB-D salient object detection (SOD) accuracy.
• Existing SOD models often struggle to balance computational efficiency with high performance.

Purpose of the Study:

• To propose an efficient RGB-D SOD model, BTNet, inspired by the primate biological visual system's P and M visual pathways.
• To improve the balance between computational efficiency and detection accuracy in RGB-D SOD.

Main Methods:

• Developed BTNet, a two-stage network simulating the M visual pathway for coarse-grained region locking and the P visual pathway for fine-grained object refinement.
• Leveraged insights from primate visual processing for distinct stage functionalities.

Main Results:

• BTNet demonstrated superior performance across six benchmark datasets compared to state-of-the-art methods.
• Achieved significant parameter reduction (93.6% less than CPNet) and high processing speed (175.4 FPS for 384x384 images).
• BTNet is nearly 7.2 times faster than the cutting-edge CPNet method.

Conclusions:

• BTNet offers an efficient and high-performance solution for RGB-D salient object detection.
• The bio-inspired design effectively balances computational cost and accuracy, outperforming existing models.
• The proposed method provides a significant advancement in efficient visual attention modeling.