SAMNet: Stereoscopically Attentive Multi-Scale Network for Lightweight Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have advanced salient object detection (SOD).
• Current CNN-based SOD methods often have large model sizes and high computational costs, limiting practical deployment, especially on mobile devices.

Purpose of the Study:

• To introduce a novel, lightweight network for salient object detection (SOD) suitable for practical, mobile applications.
• To develop a new module that enables efficient and effective feature fusion across multiple scales.

Main Methods:

• Proposed a Stereoscopically Attentive Multi-scale (SAM) module utilizing a stereoscopic attention mechanism.
• Developed SAMNet, an extremely lightweight neural network for SOD, built upon the SAM module.
• Conducted extensive experiments on popular benchmarks to evaluate performance.

Main Results:

• SAMNet achieves comparable accuracy to state-of-the-art SOD methods.
• The network demonstrates high efficiency, running at 343fps on GPU and 5fps on CPU for 336x336 inputs.
• SAMNet has a minimal parameter count of only 1.33 million.

Conclusions:

• SAMNet offers a practical and efficient solution for salient object detection (SOD).
• The proposed SAM module effectively fuses multi-scale features for lightweight SOD.
• SAMNet represents a significant step towards deploying advanced SOD systems on resource-constrained devices.