基于多任务学习的火灾检测算法

Fire Detection Algorithm Based on Multi-task Learning

烟雾和火焰是相辅相成的,在火焰检测中充分利用火焰和烟雾的内在关系,将更好地提高火灾检测的准确性。目前,大多数的火灾检测算法只关注烟雾或火焰两者间的某一个任务来设计检测算法。该文提出了一种基于多任务学习的火灾检测算法。首先,将RGB视频序列作为网络的输入,运用CNN网络与LSTM网络提取空间特征和时序特征。然后,设计生成对抗网络将特征空间划分为共享特征空间和私有特征空间两个部分,共享特征空间用于提取烟雾和火焰内在联系中的共享特征,私有特征空间用于提取烟雾和火焰各自的私有特征。最终,将上述两部分特征通过concat方式结合,送到全连接层与softmax中,得到最终的识别结果。将多任务学习应用于火灾检测,其优势在于充分考虑到火灾发生时烟雾与火焰之间的关联性,利用这种关联性丰富烟雾与火焰的特征语义信息,进而提高检测的准确性。实验结果表明,在烟雾检测的准确率上提升2.2%,火焰检测的准确率上提升1.4%,提出的多任务学习模型性能上优于单任务学习模型。

Smoke and fire are complementary to each other.Making full use of the inherent relationship between flame and smoke in flame detection will better improve the accuracy of fire detection.At present,most fire detection algorithms only focus on a certain task between smoke or flame to design detection algorithms.We propose a fire detection algorithm based on multi-task learning.Firstly,the RGB video sequence is used as the input of the network,and the CNN network and LSTM network are used to extract spatial and temporal features.Then,the generative confrontation network is designed to divide the feature space into two parts:a shared feature space and a private feature space.The shared feature space is used to extract the shared features in the internal connection of smoke and flames,and the private feature space is used to extract the private features of smoke and flames.Finally,the above two parts of features are combined by concat and sent to the fully connected layer and softmax for the final recognition result.The advantage of applying multi-task learning to fire detection is that the correlation between smoke and flame when fire occurs is fully taken into account,and the feature semantic information of smoke and flame is enriched by this correlation,thus improving the accuracy of detection.The experimental results show that the accuracy of smoke detection is increased by 2.2%,and the accuracy of flame detection is increased by 1.4%.The performance of the multi-task learning model proposed is better than that of the single-task learning model.