Context Information Based Lightweight Anti-Interference Flame Target Detection

As a kind of serious hazard,fire prevention and control has special significance in the real world.The computer vision technology is widely used in fire detection.The flame target detection problem based on deep learning is discussed,which has both theoretical significance and application value.The main contents are as follows:(1)A flame target detection method based on the improved YOLOv4-tiny is proposed.The existing flame target detection network models are large,the parameters and calculation amount are difficult to adapt to mobile devices,and some flame target detection models can meet the requirements of lightweight and real-time,but the accuracy is poor.The lightweight target detection network model is established by using the YOLOv4-tiny as the basic network and combining with the depth-separable convolution.Considering that the initial formation of flame is a small target,the multi-scale fusion scheme is proposed.Based on the fusion of shallow features,the effective channel attention mechanism is used.The experimental results on the self-built flame image data set show that compared with other existing methods,the model is more sensitive to small targets,and the real-time detection and detection accuracy are improved.(2)A hybrid flame target detection model based on Swin Transformer and YOLOv5 s is established.The Swin Transformer is introduced as a branch backbone network because convolutional neural network is difficult to use spatial context when extracting global features.As the shift window mechanism of Swin Transformer cannot adapt to the low computational requirements of mobile terminal model solving,the Ghost Module is used to deeply lightweight the branch backbone network of the YOLOv5 s.The IOU post processing algorithm is designed to solve the problem of fire-like target interference detection.The experimental results on the self-built flame image data set show that the flame target detection model and its solution method have achieved balanced improvement in real-time performance,accuracy and anti-interference.