LSNet: Lightweight Spatial Boosting Network for Detecting Salient Objects in RGB-Thermal Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current RGB-thermal salient object detection (SOD) methods are computationally intensive, hindering deployment on resource-constrained devices.
• High parameter counts and floating-point operations in existing models lead to slow inference speeds.

Purpose of the Study:

• To develop a lightweight and efficient network for RGB-thermal salient object detection (SOD).
• To improve feature extraction and reduce computational complexity for practical applications.

Main Methods:

• Proposed a lightweight spatial boosting network (LSNet) utilizing a MobileNetV2 backbone.
• Introduced a boundary boosting algorithm to optimize saliency maps and mitigate information loss.
• Employed attentive feature distillation, selection, and semantic/geometric transfer learning for enhanced feature representation.

Main Results:

• LSNet achieved state-of-the-art performance on three benchmark datasets compared to 14 existing methods.
• The model boasts significantly reduced parameters (5.39M) and floating-point operations (1.025G).
• Demonstrated high inference speeds across various hardware platforms, including common processors and GPUs.

Conclusions:

• LSNet offers an efficient and effective solution for RGB-thermal SOD, overcoming the limitations of previous methods.
• The proposed lightweight architecture and novel algorithms enable practical deployment on mobile devices.
• The study highlights the potential of optimized deep learning models for real-world salient object detection tasks.