Forest Fire Video Detection Based on Multi-scale Feature Fusion with Data Enhancement

Abstract

Due to the lack of forest fire samples and the fact that the fire images contain many targets of different sizes and complex backgrounds, overfitting is easy to occur, leading to the problems of low accuracy and high false alarm rate of forest fire detection.In this paper, by constructing the data enhancement, multi-scale feature fusion, change the activation function is put forward based on the data of enhanced multi-scale feature fusion network framework, the method through data enhancement technique to solve the problem of insufficient samples, using the deep learning automatic feature extraction fire training sample produce identification model of flames and smoke, the smoke and flames for precise identification and positioning.In addition, the activation function in the backbone network Resnet with multi-scale feature fusion was changed to Leaky Relu activation function in this paper, which was more robust to the input image noise.Through the experimental comparison and analysis of the self-built data set in this paper, it is found that the average accuracy (MAP) of forest fire detection by the network framework in this paper is increased by 14.89% and the loss value is reduced by 0.94 compared with the original data.At the same time, the fire can be monitored in real time.