摘要
针对井下安全帽检测任务中复杂环境导致检测效果不佳、模型过大难以部署的问题,通过引入ECA注意力机制改进YOLOv5s模型,在不增加计算量的前提下提高检测精度,基于Batch normalization层的模型压缩策略,裁剪网络中冗余通道,实现网络轻量化。结果表明,在自建的井下安全帽数据集中,改进方法与YOLOv5s模型检测精度相当的前提下,参数量为原网络的45.5%,有效地平衡了模型的平均检测精度和模型大小。
This paper aims to address the hard and bad detection performance of the safety helmet caused by the complex underground environments and overly large models.The study works by improving YOLOv5s model by introducing ECA attention mechanism to enhance detection accuracy with no increasing computational complexity;and cropping the redundant channels within the network to be more light-weighted by using the model compression strategy based on Batch Normalization layers.The results demonstrate that in the self-built underground safety helmet data set,the parameter is 45.5%of that of the original network,while maintaining equivalent detection accuracy between the improved method and YOLOv5s model,as which effectively balances average detection accuracy with model size.
作者
汝洪芳
梁一乐
王国新
Ru Hongfang;Liang Yile;Wang Guoxin(School of Electrical&Control Engineering,Heilongjiang University of Science&Technology,Harbin 150022,China)
出处
《黑龙江科技大学学报》
CAS
2024年第3期452-456,468,共6页
Journal of Heilongjiang University of Science And Technology
基金
黑龙江省重点研发计划指导类项目(GZ20220122)
黑龙江省省属高等学校基本科研业务费项目(2023-KYYWF-0545)。
关键词
安全帽检测
注意力机制
模型压缩
helmet detection
attention mechanism
model compression
