基于3D金字塔卷积神经网络的动态火灾预警

Dynamic fire warning based on 3D pyramid convolutional neural network

针对已有模型火灾预警实时性不足和准确率低的问题,提出一种基于3D金字塔卷积神经网络的动态火灾预警模型M。该模型是在研究3D卷积时受其结构启发构造出来的,在多帧瞬时快照上使用金字塔卷积一次性提取多尺度的特征,通过并行学习不同尺度的时序信息就能得到新的模型M。实验结果表明,模型M能够在出现火焰或烟雾后迅速进行预警,预警准确率较高。

Based on the problems of insufficient real-time performance and low accuracy of existing models for fire early warning,a dynamic fire early warning model M based on 3D pyramid convolutional neural network is proposed.The creation of the model is being inspired by its structure when studying 3D convolution.It uses pyramid convolution to extract multi-scale features at one time on multi-frame instantaneous snapshots,and a new model M can be obtained by learning time series information of different scales in parallel.The experiment results show that the model M can quickly give an early warning after the occurrence of flame or smoke and the warning accuracy is higher.