基于卷积神经网络的火灾识别算法

Fire recognition algorithm based on convolutional neural network

针对传统图像处理和浅层机器学习的火灾识别中准确率不太高、特征难以提取等问题,提出一种基于卷积神经网络的火灾识别算法。首先将图片数据集转化为快速HSI色彩格式,增加图片视觉特性,便于深度学习提取火焰特征;然后采用Inception_Resnet_V2卷积神经网络结合可变形卷积网络(DCN)对数据集进行训练提取特征,提高卷积神经网络对目标几何变化的适应和建模能力;最后使用支持向量机(SVM)分批次训练提取到的特征来进行分类。实验结果表明,与传统图像处理和其他深度学习识别算法相比,所提算法准确率高、泛化能力强、漏报率低,对测试集识别准确率达99.04%,取得很好的火灾识别效果。

In view of shortcomings of traditional image processing and shallow machine learning in fire recognition,such as difficult to extract features and low accuracy,a fire recognition algorithm based on Convolutional Neural Network(CNN)was proposed.Firstly,the picture dataset was transformed into a fast Hue,Saturation,Intensity(HSI)color format to improve image visual characteristics,so as to facilitate deep learning to extract flame features.Secondly,Inception_Resnet_V2 CNN combined with Deformable Convolutional Network(DCN)was used to train the dataset to extract features,in order to improve the adaptability and modeling ability of CNN for geometric change of target.Finally,Support Vector Machine(SVM)was used to classify the features extracted by batch training.Experimental results show that the proposed algorithm has strong generalization ability,low false alarm rate and high accuracy with 99.04%for test set in compared with traditional image processing and other deep learning recognition algorithms,and achieves good fire recognition effect.