基于全局自注意力机制的煤矸石目标检测网络
Coal gangue object detection network based on global self-attention mechanism
汝洪芳 1李作淘 1王国新 1王书侠1
作者信息
- 1. 黑龙江科技大学 电气与控制工程学院,哈尔滨 150022
- 折叠
摘要
针对在煤矸石检测过程中存在的煤与矸石相似度高而导致识别准确率低的问题,提出一种基于全局自注意力机制的煤矸石目标检测网络.将全局上下文模块引入YOLOv5s网络中,通过增加自注意力机制获得长距离特征,增大感受野,导入全维动态卷积,在控制计算量的同时进一步提升网络性能.结果表明,改进的YOLOv5s网络优于原始网络和对照组网络,与YOLOv5s、YOLOv5m、YOLOv5l、YOLOv7 和 YOLOv7x 网络相比,煤矸石目标检测精度分别提升了 4.1%、3.2%、2.1%、2.6%和2%.该模型具有较快的运行速度,满足实时检测的需求.
Abstract
This paper is focused on the solution to the problem of low recognition accuracy caused by high similarity of coal and coal gangue in the process of coal gangue detection and proposes a coal gangue object detection network based on global self-attention mechansim.The study involves introducing the global context block into YOLOv5s networ to capture long-distanced feature,enlarge receptive field,in-troduce omni-dimensional dynamic convolution by adding self-attention mechanism;and improving the network performance further as well as controling the computing load.The experimental results show that the improved YOLOv5s network is superior to the original network and the control group networks.Com-pared with the network YOLOv5s,YOLOv5m,YOLOv5l,YOLOv7 和YOLOv7x,the object detection ac-curacies of the coal gangue are improved by 4.1%,3.2%,2.1%,2.6%and 2%respectively.This model with high speed can meet the requirement of real-time detection.
关键词
煤矸石/深度学习/YOLOv5/注意力机制Key words
coal gangue/deep learning/YOLOv5/attention mechanism引用本文复制引用
基金项目
2023年黑龙江省省属高等学校基本科研业务费科研项目(2023-KYYWF-0545)
黑龙江省省重点研发计划指导类项目(GZ20220122)
出版年
2024
国家科技期刊平台
NSTL
万方数据