Fire Smoke Detection Algorithm Based On Deep Learning

Fire is the greatest threat to human life,just like floods and earthquakes.Fire and other conditions pose threats to public safety and health.It is a more frequent anomaly event than other abnormal events(such as earthquakes and floods).With the increase of fire causing factors,all kinds of fire accidents continue,which seriously threaten the safety of people’s lives and property and social harmony and stability.Obviously,extinguishing the fire can reduce the damage,but the numbers representing the area of fire and human life are still large.Therefore,we need to continuously develop implement and upgrade fire detection solutions and systems.And how to prevent fire as much as possible within our capacity is our focus.With the improvement of technology,the concept of deep learning has been introduced,and the fire smoke detection method is not limited to a few specific scenarios or sites,so the study of the fire smoke detection method has become a fire prevention technology with research value.However,false positives are often high because fire detection is complicated by natural objects that share the same characteristics as flames,dramatic changes in flame appearance,and environmental changes,such as cloud,sun,and light reflections.This paper mainly aims at the problems of low accuracy,single application scenario and difficulty in realizing real-time performance of the traditional flame and smoke detection methods mentioned above.By analyzing and studying the advantages and disadvantages of these methods,combined with my own understanding of deep learning methods and models and the characteristics of the research objectives(flame and smoke)in this paper,a fire and smoke detection algorithm based on deep learning is designed.(1)A YOLOv4-based fire smoke detection algorithm is proposed.In view of the problems of many small target samples,more overlapping and more shielding of objects to be detected in the target detection task,a new Mixup data enhancement is introduced,and is modified based on Mixup.Meanwhile,the nonlinear index moving average algorithm is used to assist the rapid convergence of the model and improve the generalization performance of the model.(2)A SEGDet-based fire smoke detection algorithm is proposed.To further investigate the problem of target detection in multiple scales and overlapping targets,a novel SEGDet model is proposed to achieve the flame and smoke detection task by implementing a semantic segmentation task combined with relevant data processing.The backbone network is similar to Segformer,which is divided into Encoder and Decoder to provide certain prior knowledge for semantic segmentation through four data enhancement methods.In the loss function,different channels can generate different detection boxes;cancel NMS algorithm and binary channels according to the results obtained by semantic segmentation,optimizing the model in overlapping target and multi-scale object detection effect.At the same time,the YOLOv4 model,the improved YOLOv4 model and the SEGDet model are implemented for the homemade smoke and flame detection data set,and the calculation effect is evaluated accordingly.The detection effectiveness of the above three models is introduced respectively.It shows that the model presented in this article is used in flame and smoke detection.