摘要
针对织物图像在电子商务、库存管理等领域的应用存在分类繁琐、检索精度不高等问题,提出了一种基于迁移学习的小样本织物图像自动分类与检索系统。首先,设计并改进了基于迁移学习的深度学习模型,对其进行微调;然后基于小样本织物图像集训练,生成新的分类模型,实现织物图像自动分类;最后,去除新模型中的分类层,提取数据集所有织物图像的图像特征,存储到Milvus向量数据库中,输入待检索织物图像,选择相似度计算方法,实现织物图像top k检索。实验结果表明:预训练模型经重新设计及训练后,织物图像识别精度可达99.5%,top 5检索的平均精度均值为0.992,平均查准率为99.65%,平均检索时间0.1653 s。通过系统的实施,可为小样本织物图像分类与检索领域现存问题提供可行的解决方案。
Aiming at the problems of cumbersome classification and low retrieval accuracy in the application of fabric images in e-commerce,inventory management and other fields,an automatic classification and retrieval system for small sample fabric images based on transfer learning was proposed.First,a deep learning model based on transfer learning was designed,improved and fine-tuned;Then,a new classification model was generated based on a small-sample fabric image set training to realize automatic classification of fabric images;Finally,the classification layer in the new model was removed,the image features of all fabric images in the dataset were extracted and stored in the Milvus vector database,the fabric images to be retrieved were input,and the similarity calculation method was selected to achieve the top k retrieval of fabric images.The experimental results show that after the redesign and training of the pre-training model,the recognition accuracy of fabric image can reach 99.5%,the average precision of top 5 retrieval is 0.992,the average precision rate is 99.65%,and the average retrieval time is 0.1653 s.Through the implementation of the system,a feasible solution can be provided for the existent problems of field of small sample fabric image classification and retrieval.
作者
游小荣
李淑芳
邓丰
雍成宇
YOU Xiaorong;LI Shufang;DENG Feng;YONG Chengyu(Changzhou Textile Garment Institute,Changzhou,Jiangsu 213164,China;Changzhou Key Laboratory of Eco-textile Technology,Changzhou,Jiangsu 213164,China)
出处
《毛纺科技》
CAS
北大核心
2023年第8期83-88,共6页
Wool Textile Journal
基金
江苏省高等学校大学生实践创新训练项目(202112807013Y)。
关键词
织物图像
迁移学习
图像分类
向量数据库
图像检索
fabric image
transfer learning
image classification
vector database
image retrieval
