YOLOv5-CSF: an improved deep convolutional neural network for flame detection

Abstract

The occurrence of fire threatens people's lives and property damage. An effective way to reduce flame damage is to detect flames in video images and take appropriate measures to contain fires at an early stage of occurrence. To solve the problem of low detection precision of current image-based flame detection methods, we propose a high precision flame detector based on yolov5 improvement. Coordinate Attention Blocks are introduced in the backbone network to obtain both channel relationships and long-range dependencies with precise positional information and increase the feature expression of the backbone network. Swin Transformer Blocks are introduced in the neck to expand the receptive field of the network and improve the flame feature extraction; the Adaptive Spatial Feature Fusion module is introduced in the head network to enhance the multi-scale feature fusion of flame features and reduce false alarms. Compared with the yolov5l, the precision of the model is improved by 4.1%. Compared with other existing detectors, it achieves the best average precision of 66.8%.