Multi-Stage Group Interaction and Cross-Domain Fusion Network for Real-Time Smoke Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Accurate smoke segmentation is crucial for fire warning systems, especially on mobile platforms.
• Existing high-precision models are often too computationally intensive for mobile applications.
• There is a need for lightweight, efficient smoke segmentation solutions.

Purpose of the Study:

• To develop a lightweight network for real-time smoke image segmentation on mobile devices.
• To improve the analysis of smoke features and preserve information from small smoke objects.
• To enhance the boundary detection of smoke targets.

Main Methods:

• Proposed a Multi-stage Group Interaction and Cross-domain Fusion Network (MGICFN) with low computational complexity.
• Introduced a Cross-domain Interaction Attention Module (CIAM) for merging spatial and frequency domain features.
• Designed a Multi-stage Group Interaction Module (MGIM) to prevent information loss during downsampling and an Edge Enhancement Module (EEM) for boundary refinement.

Main Results:

• MGICFN achieved 88.70% Dice and 81.16% mIoU on the SFS3K dataset.
• The model obtained 87.30% Dice and 78.68% mIoU on the SYN70K test dataset.
• The MGICFN model has only 0.73M parameters and 0.3G FLOPs, indicating high efficiency.

Conclusions:

• The proposed MGICFN is an effective and computationally efficient solution for lightweight smoke image segmentation.
• The network demonstrates strong performance on benchmark datasets, suitable for real-time mobile applications.
• MGICFN offers a promising approach for enhancing fire warning systems through improved smoke detection.