Fire Detection Algorithm Based on Improved YOLOv5

Abstract

Aiming at the problems of low accuracy and slow speeding of manual and sensor detection in existing farm fire detection. This paper proposed a real-time fire detection algorithm based on improved yolov5s. The algorithm based on YOLOv5s network. First, the SE channel attention mechanism was improved by fusing maximum pooling and average pooling to enhance the receptive field, and improved the detection accuracy of small targets and blurred smoke boundaries. Secondly, Ghost model was used to reduce the dimension of the volume layer of the target recognition network, in order to reduce the number of model parameters and convolution computation, thus improve the detection rate.Experiments showed that the improved yolov5s real-time fire detection model can automatically identify and detect flames, smoke in different stages and forms. Finally, Automatic smoke, flame recognition and detection were realized. The detection results of average mAP@0.5 value was 80.6%, which improved 3.2% compared with the original model. And the targets detection results of speeding was 31FPS. The test results showed that the real-time fire detection models of different combustion stages and different forms were improved. Improved model had higher detection accuracy, the calculation speed was faster, and the timeliness was stronger, which improved the efficiency of fire detection, and reduced the loss caused by insufficient manpower and the detection effect of sensors, improved work efficiency.