YOLOFM: an improved fire and smoke object detection algorithm based on YOLOv5n.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing fire detection algorithms struggle with feature extraction, computational complexity, and accuracy, limiting their deployment on resource-constrained devices.
• Challenges include missed and inaccurate detections, necessitating more robust and efficient solutions.

Purpose of the Study:

• To develop a highly accurate and efficient fire detection algorithm, YOLOFM, addressing limitations of current methods.
• To improve multi-scale information integration, reduce model parameters, and minimize redundant calculations for better performance.

Main Methods:

• Developed YOLOFM algorithm using a FocalNext network with FocalNextBlock and a novel QAHARep-FPN architecture.
• Introduced a new compression decoupled head (NADH) and proposed Focal-SIoU loss for bounding box regression.
• Manually labeled a dataset of 18644 images (FM-VOC Dataset18644) using LabelImg software.

Main Results:

• YOLOFM demonstrated significant improvements over the baseline network: 3.1% in accuracy, 3.9% in recall, 3.0% in F1-score, 2.2% in mAP50, and 7.9% in mAP50-95.
• The algorithm achieved a balance between performance and speed, offering a dependable solution for fire detection tasks.

Conclusions:

• YOLOFM effectively overcomes limitations in current fire detection algorithms, providing enhanced accuracy and efficiency.
• The proposed methods, including FocalNext network, QAHARep-FPN, NADH head, and Focal-SIoU loss, contribute to a more robust and deployable fire detection system.