基于多分辨率卷积神经网络的火焰检测

Fire Detection Based on Multi-resolution Convolution Neural Network

采用一种基于多分辨率卷积神经网络的火焰检测算法对真实场景中的火焰目标进行检测.该算法以BN_Inception网络为基础架构,采用不同分辨率的神经网络互补学习复杂场景中火焰的多尺度视觉特征,同时该算法重点关注检测目标场景的背景环境、局部目标和整体布局等特征.使用该算法在构造的涵盖大多数真实场景的火焰数据集上进行测试,实验结果表明,提出的算法能够取得更好的检测效果,并在实际场景中得到了有效验证.

A fire detection algorithm based on multi-resolution convolutional neural network was proposed to achieve the objective of fire detection in real scenes. This algorithm leveraged the BN_Inception network as the basic network structure. Different coarse and fine resolution neural networks were used to learn the multi-scale visual features of the fire in complex scenes complementarily, while paying attention to the background environment, local targets and overall layout of the scene. The method was evaluated in fire dataset covers most real scenes. The experiment showed that the proposed method could achieve better detection results than other methods, which could be effectively applied in the real world.