| 摘 要: 针对火焰检测参数量和计算量较大及准确度较低的问题,提出一种基于Fire-MCANet(Fire-Max Convolution Activate Networks)的火焰检测模型。该模型首先构建一种MCA(Max Convolution Activate)模块,使用大卷积核获取感受野,提高特征提取的能力;其次构建主干网络MCANet Block,在提升感受野的同时,降低模型的参数量和计算量;最后引入CA(Coordinate Attention)注意力机制获取火焰的位置信息。实验结果表明,基于Fire-MCANet的火焰模型的检测准确率达到95.75%,计算量仅有2.13 GMac;其网络模型的结构比ConvNeXt网络更加轻量化,检测效果也更好。 |
| 关键词: 火焰检测;深度学习;CA注意力机制;特征提取 |
|
中图分类号: TP391
文献标识码: A
|
|
| Flame Detection Model Based on Fire-MCANet |
|
ZHU Qiaoqiao, YAN Yunyang, LENG Zhichao, DONG Ke, YE Xiang, WANG Panlong
|
(Faculty of Computer & So f tware Engineering, Huaiyin Institute of Technology, Huaian 223003, China)
1546906478@qq.com; yunyang@hyit.edu.cn; 1067321462@qq.com; 1553789590@qq.com; 1528411799@qq.com; 905610658@qq.com
|
| Abstract: Aiming at low accuracy of flame detection with large number of parameters and calculations, this paper proposes a flame detection model based on Fire-MCANet (Fire-Max Convolution Activate Networks). Firstly, a Max Convolution Activate (MCA) module is constructed to obtain the receptive field by using a large convolutional kernel to improve the ability of feature extraction. Secondly, the backbone network MCANet Block is constructed to improve the receptive field and reduce the number of parameters and calculations of the model. Finally, the CA (Coordinate Attention) attention mechanism is introduced to obtain the position information of the flame. The experimental results show that the detection accuracy of the flame detection model based on Fire-MCANet reaches 95.75% , and the computational amount is only 2.13 GMac. Its network model is lighter than the ConvNeXt network, and the detection effect is better. |
| Keywords: flame detection; deep learning; CA attention mechanism; feature extraction |
