Research And Design Of Forest Fire Recognition System Based On Convolutional Neural Network

In recent years,forest fires have occurred frequently around the world.Once a forest fire starts to spread,it is very difficult to extinguish it in a short period of time and will cause very large losses,so forest fires need to be identified.The identification of forest fires is a very complex task due to the different shapes,textures and colors of forest fires.There are two main approaches for forest fire recognition.The first approach uses traditional image processing methods to identify forest fires,while the second approach uses deep learning algorithms to identify forest fires.Traditional image processing methods require manually designed feature extraction algorithms based on priori knowledges,and the manually designed feature extraction algorithms often have insufficient generalization capability because flames at different stages have different features.Convolutional neural network(CNN)in deep learning algorithms can automatically extract deep features in images,avoiding the complexity and blindness of the feature extraction stage.In order to improve the reliability of forest fire monitoring,this paper uses deep learning technology to identify forest fires,deploys deep learning models on embedded hardware,and combines sensor technology with Io T technology to design a monitoring system for forest fires.The research of this paper consists of the following parts:(1)An image dataset of forest fires was constructed.First,an image dataset of forest fires was constructed using a Google search engine to select high-quality images,and the dataset was divided into a training set and a test set at a certain ratio.Secondly,in order to solve the problem of insufficient forest fire image dataset,the training set images are PCA＿Jittering processed to expand the training set by drastically changing its main colour palette while ensuring that the overall hue of the training set images remains unchanged.Finally,to reduce noise and fog interference in the image,the image is filtered using Gaussian filtering and the image defogging algorithm is switched on when there is fog in the forest.(2)A CNN-based forest fire recognition model is constructed and a segmented training method is proposed.First,the CNN model is constructed,the residual structure is added to the model and the fully connected layer of the model is modified,while transfer learning is used for training in order to reduce overfitting;Secondly,to further reduce model overfitting,a segmented training method is proposed,which dynamically adjusts the resolution of the input image during model training,learning and extending the overall structure of the image;Finally,to verify the performance of the constructed model,the CNN was compared with the traditional forest fire identification model.(3)In order to further improve the accuracy of the CNN model and reduce the latency time,an improved two-channel CNN is proposed and compared with different models.Firstly,in order to adapt to forest fire scenes of different sizes,a dual-channel CNN is proposed,which is trained simultaneously and extracts features independently for fusion.Secondly,the fully connected layer of the two-channel CNN was replaced with a support vector machine(SVM),taking into account the unique advantages of SVM for small samples and classification problems.Then,in order to reduce the delay time,the improved two-channel CNN is proposed by adding a feature compression layer to the structure of the two-channel CNN and designing Lasso＿SVM to replace the SVM in the original model.Finally,the improved dual-channel CNN was compared and analysed with other models,and it was found that the improved dual-channel CNN achieved98.475% recognition accuracy for forest fires,with an average delay time of 0.051 sec/frame,proving the advantages of the model in terms of high accuracy and low delay.(4)An embedded platform for forest fire monitoring systems was built.First,deployment of a modified two-channel CNN on an embedded hardware Raspberry Pi 4B for forest fire identification;Secondly,to further improve the reliability of forest fire monitoring,different types of sensors were selected to receive data from the forest environment;Then,in order to visualise the data and improve data utilisation,the forest fire identification results and forest environment data were uploaded via Zig Bee to the host computer and the One Net Io T platform;Finally,to further reduce the false alarm rate of forest fires,the forest images captured by the Raspberry Pi 4B are transmitted to the host computer in real time for review of forest images at forest fire monitoring points.The innovation of this paper lies in the following three main points:(1)First,this paper improves the CNN model by using transfer learning and proposing a segmented training method to reduce the overfitting of the model;Secondly,a two-channel CNN model is proposed in this paper based on an improved CNN model;Finally,this paper proposes the Lasso＿SVM layer to compress the features extracted by the two-channel model to obtain an improved two-channel CNN,which achieves high accuracy and low latency in forest fire identification;(2)In this paper,the improved dual-channel CNN is deployed on an embedded platform to complete the recognition of forest fires,while sensor technology and Io T technology are combined to the platform to achieve the transmission and sharing of forest environment data and image recognition results,further reducing the false alarm rate of forest fires.