摘要
针对现有火焰识别算法在效率上的不足,设计1种轻量高效的深度学习模型。模型基于SqueezeNet进行优化,引入双分支注意力机制以强化对火焰特征的识别能力,提升模型分类性能;同时,加入残差连接,提高网络的训练稳定性和特征表达能力;通过使用批通道归一化技术提高网络的泛化性能;此外,通过将Fire模块中的3×3标准卷积核替换为深度可分离卷积,进一步降低参数数量和计算复杂度,并通过多个公开的火焰图像数据集来评估所提算法的性能。研究结果表明:相较于原始的SqueezeNet算法,改进后的SqueezeNet模型不仅提升检测速度,同时也获得更高的识别准确率和更好的泛化能力。研究结果可为实时火灾监测系统和智能消防设备的开发提供理论基础和技术支持。
Addressing the efficiency limitations of existing flame recognition algorithms,a lightweight and efficient deep learning model was designed.The model was optimized based on SqueezeNet,then the dual-branch attention mechanism was introduced to enhance the recognition capability of flame features,and the classification performance of the model was improved.Meanwhile,the residual connections were incorporated to increase the training stability and feature representation capacity of network.The batch channel normalization technology was employed to enhance the generalization performance of the model.Furthermore,the 3×3 standard convolution kernels in the Fire module were replaced with depth-wise separable convolutions,further reducing the parameter amount and computational complexity.The performance of the algorithm was evaluated using multiple public flame image datasets.The results show that compared to the original SqueezeNet algorithm,the improved SqueezeNet model not only enhances the detection speed,but also achieves higher recognition accuracy and better generalization ability.This research results can provide theoretical foundation and technical support for the development of real-time fire monitoring systems and intelligent firefighting equipment.
作者
王文标
时启衡
郝友维
WANG Wenbiao;SHI Qiheng;HAO Youwei(College of Marine Electrical Engineering,Dalian Maritime University,Dalian Liaoning 116026,China)
出处
《中国安全生产科学技术》
CAS
CSCD
北大核心
2024年第8期19-26,共8页
Journal of Safety Science and Technology
基金
国家自然科学基金项目(52071047,62073054)。
