FM-Net: A New Method for Detecting Smoke and Flames

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Existing smoke and fire detection algorithms struggle with high false alarm rates and poor interference resistance in complex scenes.
• Traditional convolutional operations can lead to edge feature loss during down-sampling, hindering accurate detection.

Purpose of the Study:

• To develop an advanced target detection network for smoke and fire detection with improved accuracy and robustness.
• To address the limitations of existing algorithms in complex environments and real-time applications.

Main Methods:

• Proposed a novel target detection network utilizing an improved feature pyramid structure.
• Introduced Context Guided Convolutional Block to fuse target and environmental information, mitigating edge feature loss.
• Designed Poly Kernel Inception Block for multi-scale feature extraction and collaborative characterization of flame and smoke.
• Incorporated Manhattan Attention Mechanism Unit for enhanced spatial-temporal correlation and pixel-level dependency modeling.

Main Results:

• The proposed algorithm demonstrated significant advantages in detection accuracy compared to existing lightweight models.
• Experimental results on a self-constructed dataset validated the algorithm's effectiveness.
• The system satisfies the demand for real-time smoke and fire detection.

Conclusions:

• The improved feature pyramid network effectively enhances smoke and fire detection accuracy and interference resistance.
• The novel network components contribute to better feature fusion, multi-scale analysis, and spatial-temporal correlation.
• This algorithm presents a promising solution for reliable and real-time fire detection systems.