基于多尺度空洞卷积自编码神经网络的森林烟火监测

Forest Fire Detection Based on Multi-Scale Dilated Convolution Auto-Encoding Neural Network

森林火灾范围大距离远,火灾图像中有效特征提取尺寸较大,传统卷积网络难以有效学习,另外火灾中烟雾和雾气较为相似,容易造成错误识别。针对上述问题,提出一种基于多尺度空洞卷积自编码器(Multi-Scale Dilated Convolution Auto Encoder,MSDCAE)的深度网络,通过空洞卷积获得不同尺寸的感受野特征并连接输出来优化特征学习,再基于Softmaxwithloss设计改进的损失函数(Improved Softmaxwithloss,ISWL)来提升烟雾、雾气等相似图像的分类性能。实验验证了MSDCAE自编码器和ISWL损失函数的有效性,结果证明在森林火灾的烟火图像识别中,新方法对比普通深度网络算法更具优越性。

Large range of forest fires cause the large size of effective feature in fire images,which are difficult to be learned effectively by traditional convolution networks.In addition,because fire smoke and fog are similar,it is easily to recognize error.Aiming at these problems,a new deep network based on weighted multi-scale dilated convolution auto-encoder(MSDCAE)is proposed.Different sizes of receptive field features are obtained through different dilated convolution kernels and concatenated output to optimize feature learning.Based on softmaxwithloss,improved loss function(ISWL)is designed to improve the classification performance of similar images such as fire smoke and fog.Experiments verify the effectiveness of MSDCAE auto-encoder and ISWL loss function.The results prove that the new method is superior to the ordinary deep network algorithm in the image recognition of forest fires.