An efficient fire detection algorithm based on Mamba space state linear attention.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Object detection models often struggle with complex scale variations and multi-view interference in dynamic environments.
• State Space Models (SSMs) offer enhanced global receptive fields and feature extraction compared to traditional methods.
• Existing visual fire detection technologies require performance enhancements for real-world applications.

Purpose of the Study:

• To propose an efficient and high-performance fire detection algorithm by integrating the Mamba attention mechanism into the YOLOv9 architecture.
• To introduce novel techniques for improving feature representation, local detail modeling, and bounding box accuracy in fire detection.
• To provide a computationally efficient solution for visual fire detection.

Main Methods:

• Developed the Efficient Mamba Attention (EMA) module, combining adaptive average pooling with SSMs to reduce computation and enhance feature representation.
• Integrated the ConvNeXtV2 module into the backbone network to improve the modeling of fine-grained local features.
• Refined the loss function with a dynamic non-monotonic focusing mechanism and distance penalty strategy for better bounding box accuracy.

Main Results:

• The proposed model achieved an FPS of 71.
• Achieved a mean Average Precision ([Formula: see text]) of 91.0% on a large-scale fire dataset and 87.2% on a small-scale fire dataset.
• Demonstrated superior performance and significant computational efficiency compared to existing fire detection methods.

Conclusions:

• The Mamba-enhanced YOLOv9 model offers a powerful and efficient solution for visual fire detection.
• The novel EMA module and backbone optimizations significantly improve detection accuracy and robustness.
• This approach represents a significant advancement in real-time fire detection systems for dynamic environments.