A Lightweight Cross-Layer Smoke-Aware Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Smoke detection is crucial for pre-fire identification.
• Existing methods struggle with variable smoke morphology, limiting practical applications.

Purpose of the Study:

• To propose a lightweight cross-layer smoke-aware network (CLSANet) for precise smoke characteristic extraction.
• To enhance information exchange and feature extraction accuracy in smoke detection.

Main Methods:

• Developed a lightweight CLSANet (2.38 M) with three cross-layer connection strategies.
• Introduced a spatial perception module (SPM) for shallow-to-deep layer information transfer.
• Designed a texture federation module (TFM) using fully connected attention (FCA) and spatial texture attention (STA) for spatial information repair.
• Implemented a feature self-collaboration head (FSCHead) to decouple localization and classification tasks.

Main Results:

• CLSANet achieved 94.4% precision on the USTC-RF database and 73.3% on the XJTU-RS database.
• The network effectively removes redundancy while preserving meaningful smoke features.
• Extensive experiments confirmed CLSANet's competitive performance in smoke detection.

Conclusions:

• CLSANet offers an effective and concise solution for precise smoke feature extraction.
• The proposed cross-layer strategies significantly improve smoke detection accuracy.
• CLSANet demonstrates strong potential for practical pre-fire detection applications.