采用Anchor-Free网络结构的实时火灾检测算法

为了解决现有的火灾检测算法中模型复杂，实时性差，检测精度较低的问题，提出快速高效的火灾检测算法. 该算法采用Anchor-Free网络结构，克服了Anchor方法中超参数过多、网络结构复杂的缺点；选用MobileNetV2作为基础特征提取网络，满足了检测的高实时性需求；针对火焰目标的形状、尺度多变的特点，设计适合于火焰检测的分类与边框预测子网络；引入特征选择模块，在特征金字塔网络中自动选择更合适的金字塔特征层. 算法在自建数据集上的检测精度达到90.1%；在公开数据集上取得了较好的检测结果，其检测速度可达24.6 帧/s. 实验结果表明，算法的网络模型简单，检测精度较高，检测速度较快；综合性能优于现有的其他火灾检测算法.

Real-time fire detection algorithm with Anchor-Free network architecture

The existing fire detection algorithms had problems such as too complicated model structure, poor real-time ability, and low detection rate. A fast and efficient fire detection algorithm based on convolutional neural network was proposed, in order to solve those problems. The Anchor-Free network structure was adoped as the whole architecture, which avoided the shortcomings of too many hyper-parameters and complex network structure in the Anchor-based detector. MobileNetV2 was selected as the feature extraction backbone network in order to meet the requirements of high real-time ability. A classification and a box prediction task-specific subnet suitable for fire detection were designed considering that the size and shape of the flame were different from other objects. A feature selective module was introduced to help the feature pyramid network automatically selected a more suitable pyramid feature layer. The algorithm achieves 90.1% accuracy on the self-built dataset, and has good performance on the public fire dataset. The speed of the network achieves 24.6 frames per second. The experimental results show that the detection precision of the proposed algorithm is high and the speed is fast while being simpler. The comprehensive performance of the algorithm is better than other existing fire detection algorithms.