Research and optimization of a multilevel fire detection framework based on deep learning and classical pattern...

Area of Science:

• Computer Science
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Traditional fire detection methods and existing deep learning models face challenges with accuracy and false alarms in complex environments.
• Effective flame detection is crucial for public safety and property protection.

Purpose of the Study:

• To introduce the Fire Focused Detection Network (FFDNet), a novel flame detection framework.
• To enhance flame detection sensitivity, precision, and reduce false alarm rates.

Main Methods:

• Integration of an enhanced Real-Time DEtection TRansformer (RT-DETR) model with Vector Quantized Generative Adversarial Network (VQGAN).
• Incorporation of a novel loss function, Innovative Minimum Perimeter Distance IoU (InnMPD-IoU), into the RT-DETR model.
• Utilizing Complete Local Binary Pattern (CLBP) for texture feature extraction and VQGAN for flame identification via sample reconstruction.

Main Results:

• FFDNet achieved high performance metrics on the Dataset for Fire and Smoke Detection (DFS): 98.23% precision, 96.33% recall, 97.33% F1 score, and 95.08% accuracy.
• The proposed method significantly surpasses existing fire detection techniques.
• Demonstrated enhanced sensitivity and reduced false alarm frequencies.

Conclusions:

• FFDNet represents a state-of-the-art flame detection framework with superior performance.
• The developed methodology shows potential for robust and efficient flame detection applications.
• FFDNet offers significant technical support for fire prevention and response initiatives.