Instance segmentation using semi-supervised learning for fire recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Processing

Background:

• Fire disasters pose significant risks to life and property, necessitating timely identification via image processing.
• Current fire instance segmentation methods struggle with limited data, low accuracy, and slow inference speeds.

Purpose of the Study:

• To develop an accurate and efficient fire instance segmentation method using semi-supervised deep learning.
• To address challenges of data scarcity and improve recognition accuracy and inference speed in fire detection.

Main Methods:

• Utilized a lightweight SOLOv2 network with optimized structure for fire instance segmentation.
• Implemented a novel semi-supervised learning approach incorporating fire feature matching (color, morphology) for pseudo-label refinement.
• Employed strong image enhancement techniques for selected images to improve student model training.

Main Results:

• The proposed semi-supervised method significantly enhances fire instance segmentation accuracy.
• The optimized network and training strategy maintain efficient inference speeds.
• Demonstrated improved performance even with limited datasets.

Conclusions:

• The developed semi-supervised deep learning method offers a promising solution for accurate and efficient fire instance segmentation.
• The approach effectively mitigates issues related to data limitations and pseudo-label errors.
• This technique contributes to improved fire disaster response through advanced image processing.