Forest Fire Image Recognition Algorithm And Realization Based On Deep Learning

The occurrence of forest fire has a great impact on the economy, the ecological environment and so on. For this reason, study on forest fire monitoring system is very concerned by world countries. Sensor recognition is greatly influenced by the environment, which is not suitable for application on forest fire monitoring in large scale. Traditional image methods need to be processed on the image, artificial feature extraction is necessary. Feature selection has become the key factor to achieve the desired results. Artificial feature extraction is a waste of time and energy. For the above, this paper use Deep Learning algorithms which have developed rapidly in the field of machine learning. The model of Convolutional Neural Network was applied in forest fire recognition. Deep network can automatically extract the features of input image, and bottom layer features can be combined to form high layer features by passing between layers. It can avoid the complexity and blindness of artificial feature extraction in traditional method. Local receptive field and weight sharing technology reduce the number of parameters and the complexity of the algorithm of convolutional neural network so that the difficulty of algorithm training greatly reduce. Subsampling can tolerate the image distortion to some extent and this enhances the robustness of network. Experiments showed that this method has achieved satisfactory results.The main work of this article is as follows:(1) A forest fire database was created through experiments and online collection. (2) The research extracted image features manually and used the commonly methods such as the Support Vector Machine, the Radial Basis Function network and the Backward Propagation neural network to identify the features and analyze the results. (3) The research identified the forest fire image has fire or not based on the depth research of Convolutional Neural Network. For the different backgrounds between day and night, different networks were designed to identify. Furthermore, different structures and parameters were compared and analyzed. The accuracy of night forest fire recognition model was 95.71%, the accuracy of day forest fire recognition model was 98%. It had significant advantages compared with traditional image methods.