Design Of Forest Fire Detection System Based On MobileNet-SSD

In recent years,forest fires happen frequently,which will destroy the original ecological balance of the forest,and at the same time bring a lot of property losses and even kill human life.Therefore,rapid and accurate detection and early warning in the early stage of forest fire can reduce the loss value,which is of great significance for the protection of forest resources.The forest fire detection system based on traditional image processing is prone to misinformation and missing information in the complex and changeable forest environment.This paper designs a forest fire detection system based on lightweight network model by studying and analyzing deep learning related algorithms.In order to improve the detection speed and accuracy,it mainly includes the following contents:The SSD(Single Shot Multi Box Detector)algorithm is selected as the basic framework network of the forest fire detection system in this paper.It also absorbs the excellent design ideas of the Faster R-CNN and YOLO algorithms,and adopts the characteristic maps of different sizes to predict the target objects to be detected.First of all,aiming at the problem that the complexity of the feature extraction network in SSD algorithm causes a large amount of computation,the network structure needs to be lightweight improved.In the commonly used lightweight network model,MobileNet network has better performance than Squeeze Net and ShuffleNet networks on FLAME and D-Fire forest fire data sets due to the use of deep separable convolution.In this paper,the backbone network of SSD is replaced by VGG16 with MobileNet network,and a forest fire detection algorithm based on MobileNet-SSD is constructed.Secondly,in view of the problem that the MobileNet-SSD forest fire detection algorithm is not sensitive to the detection of small target forest fire,this paper further integrates void convolution and SENet network into the model at the same time.On the one hand,the hole convolution with different expansion rates is introduced into the shallow feature map of the algorithm to expand the receptive field of features and improve the semantic information of small targets.On the other hand,the SENet channel attention network module is introduced into the high-level feature map of the algorithm to improve the ability of the model to extract key features of medium and large targets.In the process of model training,cross entropy loss function and Adam optimization algorithm are designed to make the model converge quickly.In addition,L2 regularization structure is added to the MobileNet network to prevent over-fitting problems.The final experiment shows that the detection accuracy of the improved model on the self-built forest fire test data set is up to 94.75%,which is 6.5% higher than the unmodified model,and can effectively identify the forest fire areas in different environments.In addition,during the test,the whole system functions normally and can provide fire location and fire information in case of forest fire.