摘要
为了提高纺织行业的织物管理效率,解决织物检索耗时久、检索精度低等问题,使用改进的深度学习网络LResNet50E⁃IR得到织物图像的特征表示,利用faiss索引进行特征相似度匹配后,排序输出检索的结果。同时,提出分级检索策略,根据不同类别的织物特点先分类,再通过改进的神经网络训练得到相应检索模型,最后查询未入数据库的织物图像,来验证检索的效果。试验结果表明:本系统检索的top10 mAP高达99.22%,检索速度快。认为:该织物图像检索方法可以满足准确性与高效性的要求。
In order to increase the fabric management efficiency of textile industry,solve the problems like long time-consuming of fabric retrieval,lower retrieval accuracy and so on,the improved deep learning network LResNet50E-IR was used to obtain the character representation of fabric images.Faiss index was used to match similarity of characteristics,then the results were sorted and output.Meanwhile,classification retrieval strategy was put forward.According to fabric characteristics of different classification,sorting was firstly done,then corresponding retrieval model was obtained by improved neural network training.Finally,by searching the fabric images without imputing in database,the retrieval effect was verified.The experimental results showed that top 10 mAP retrieved by the system was reached up to 99.22%.The retrieval was fast.It is considered that the fabric image retrieval method can meet the requirements on accuracy and high efficiency.
作者
刘瑞昊
于振中
孙强
LIU Ruihao;YU Zhenzhong;SUN Qiang(Jiangnan University,Wuxi,214122,China;HRG International Institute(Hefei)of Research and Innovation,Hefei,230601,China)
出处
《棉纺织技术》
CAS
北大核心
2022年第5期42-47,共6页
Cotton Textile Technology
基金
安徽省科技攻关计划(202003a05020015)。
关键词
织物
图像检索
深度学习
卷积神经网络
分级检索
faiss索引
fabric
image retrieval
deep learning
convolutional neural networks
classification retrieval
faiss index
