Lightweight Salient Object Detection via Hierarchical Visual Perception Learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) have advanced Salient Object Detection (SOD).
• Current high-accuracy SOD models are computationally intensive and large, hindering deployment on mobile devices.
• There is a need for efficient SOD models suitable for real-world applications.

Purpose of the Study:

• To develop a lightweight Salient Object Detection (SOD) model for practical applications, especially on mobile devices.
• To mimic the hierarchical visual processing of the primate visual system for improved perception learning.
• To design and evaluate a novel network architecture for efficient SOD.

Main Methods:

• Proposed a Hierarchical Visual Perception (HVP) module inspired by primate visual cortex processing.
• Developed HVPNet, a lightweight SOD network incorporating the HVP module.
• Conducted extensive experiments on popular benchmarks to evaluate performance.

Main Results:

• HVPNet achieves highly competitive accuracy compared to state-of-the-art SOD methods.
• The model demonstrates efficient performance with only 1.23M parameters.
• Achieved 4.3 frames/s CPU speed and 333.2 frames/s GPU speed.

Conclusions:

• HVPNet offers a promising solution for deploying accurate SOD on resource-constrained platforms.
• The hierarchical visual perception approach effectively enhances SOD performance in a lightweight architecture.
• The developed model balances accuracy and efficiency for real-world SOD applications.