基于改进YOLOv7的快速火灾检测算法分析

Analysis of High-speed Fire Detection Algorithm Based on improved YOLOv7

阐述基于改进YOLOv7的火灾检测算法。首先通过将YOLOv7的主干网络替换为轻量级网络MobileOne,把一个k×k的卷积层等价转换为多个分支,使得网络在训练时参数更多、学习能力更强,从而在推理阶段具有更高的计算速度。接着在池化层引入SPPFCSPC,在特征提取时获得不同的感受野,有效地解决目标差异较大的问题,同时该结构相比于原SPPCSPC计算量减少一半,节约计算成本。最后利用模型剪枝等压缩方法对其做轻量化处理,让模型推理速度进一步提升。

This paper describes an improved fire detection algorithm based on YOLOv7.Firstly,the backbone network of YOLOv7 was replaced through the lightweight network MobileOne,which converts a k×k convolution layer into multiple branches equivalent to each other,which makes the network more complex in training,more parameters,stronger learning ability,but higher computing speed in the reasoning stage.Secondly,SPPFCSPC was introduced in the pooling layer to obtain different receptive fields during feature extraction,which effectively solves the situation of large difference in objectives.Meanwhile,compared with the original SPPCSPC,the calculation amount of this structure was reduced by half and the calculation cost was saved.Finally,model pruning and other model compression methods were used to carry out lightweight processing to further improve the reasoning speed of the model.