Real-time factory smoke detection based on two-stage relation-guided algorithm.

Area of Science:

• Computer Vision
• Environmental Monitoring
• Artificial Intelligence

Background:

• Image-based air quality analysis faces challenges due to variations in real-world smoke images.
• Existing object detection methods struggle with the complexities of smoke detection.

Purpose of the Study:

• To propose an efficient, mobile-platform-compatible smoke detection algorithm for real-time factory supervision.
• To enhance smoke detection accuracy and speed using image analysis techniques.

Main Methods:

• Developed a Two-Stage Smoke Detection (TSSD) algorithm utilizing a lightweight detection framework.
• Integrated prior knowledge and contextual information into a relation-guided module to refine the smoke search space.
• Ensured compatibility with various image resolutions.

Main Results:

• The TSSD algorithm significantly improved the accuracy of single-stage detection methods.
• Achieved a detection accuracy of 59.24%, outperforming Faster RCNN.
• Demonstrated a real-time detection speed of 50 ms (20 FPS).

Conclusions:

• The TSSD algorithm offers high stability, accuracy, and speed for smoke detection.
• The system is suitable for mobile terminals and various environmental monitoring applications.
• Provides a robust solution for real-time smoke detection and supervision.