摘要
提出了一种基于改进YOLOv5s网络模型的火灾图像识别方法。通过引入注意力机制改进特征提取网络,提高模型对特征的学习能力;通过添加大尺度检测层改进多尺度检测机制,执行K-Means聚类算法改进先验框,增强模型对小目标的识别能力。在实验数据集上的测试结果表明:改进的YOLOv5s网络模型相比原始模型在精确率、召回率和平均精度均值(mAP)指标上均有提升。改进模型的mAP为85.72%,帧率达54.66fps;在置信度上有了明显提升,对多目标和小目标的识别效果更好,并且有效降低了漏检和误检情况。所提出的火灾图像识别方法可适用于安防监控系统或智能机器人。
A fire image recognition method based on improved YOLOv5s network model is proposed.Feature extraction network is improved by introducing attention mechanism to improve the learning ability of the model to learn about feature.Multi-scale detection mechanism is improved by adding large-scale detection layer,and K-Means clustering algorithm is implemented to improve the anchor box,which enhances recognition ability of the model for small targets.Test results on the experimental dataset show that the improved YOLOv5s network model has improved in terms of precision,recall and mean average precision(mAP)comparing with the original model.The mAP of improved model is 85.72%,and the frame rate reaches 54.66 fps.The improved model has a significant improvement in the level of confidence,the recognition effect on multiple targets and small targets is better,and the missed and false detections are effectively reduced.The proposed fire image recognition method is applicable for security monitoring systems and intelligent robots.
作者
梁金幸
赵鉴福
周亚同
史宝军
LIANG Jinxing;ZHAO Jianfu;ZHOU Yatong;SHI Baojun(State Key Laboratory of Reliability and Intelligence of Electrical Equipment,Hebei University of Technology,Tianjin 300401,China;School of Mechanical Engineering,Hebei University of Technology,Hebei Key Laboratory of Robot Sensing and Human-Robot Integration,Tianjin 300401,China;School of Electronic Information Engineering,Hebei University of Technology,Tianjin 300401,China)
出处
《传感器与微系统》
CSCD
北大核心
2024年第1期157-161,共5页
Transducer and Microsystem Technologies
基金
国家重点研发计划资助项目(2019YFB1312102)
国家自然科学基金资助项目(U20A20201)
河北省重点研发计划资助项目(20311803D)
河北省自然科学基金资助项目(E2019202338)。
