基于细粒度特征的面料图像检索
Fabric Image Retrieval Based on Fine-Grained Features
罗辛 1夏冬梅 1陶然 1史有群1
作者信息
- 1. 东华大学 计算机科学与技术学院,上海 201620
- 折叠
摘要
面料图像检索对于纺织工厂面料库存和样品管理意义重大,但面料外观的多样性以及织物纹理的精细性,使得在面料检索时面料的特征提取较困难.该研究提出一种基于细粒度特征的面料图像检索算法.该算法使用坐标注意(coordinate attention,CA)模块来提取图像的精准位置信息,并将缩放系数法用于在宽度和高度方面整体缩放MobileNetV3 的网络结构以减少模型参数数量,达到减少网络训练时间的目的.据此筛选出提取面料图像细粒度特征的最佳模型,在面料图像数据集(fabric image dataset,FID)上进行面料检索实验.结果表明,该算法有效提高了面料图像细粒度特征提取的准确性,检索精度达到 91.82%,浮点运算数达到175.34 MB.检索精度比MobileNetV3 原模型提高了 13.49 个百分点,同时减少了网络训练时间,速度提高了25.14%.该算法具有实际应用价值.
Abstract
Fabric image retrieval is crucial for textile mills to manage their inventory and samples,but it is challenging due to the diverse appearance and fine-grained texture of fabrics.This paper proposes an algorithm based on fine-grained features to deal with this issue.The algorithm uses the coordinate attention(CA)module to extract precise location information of the fabric images and scales the overall network structure of MobileNetV3 to reduce the training time and model parameters.The optimized model is selected based on the scaling factor method,and fabric retrieval experiments are conducted on the fabric image dataset(FID).The results show that the algorithm effectively improves the accuracy of fabric image feature extraction,with a retrieval accuracy(Acc)of 91.82%and floating point operations(FLOPs)of 175.34 MB.The Acc is improved by 13.49 percentage points compared with that of the original MobileNetV3 model,while the training time is reduced,and the inference speed is improved by 25.14%.The algorithm has practical application value.
关键词
面料图像检索/MobileNetV3/细粒度特征/注意力机制/缩放系数Key words
fabric image retrieval/MobileNetV3/fine-grained feature/attention mechanism/scaling factor引用本文复制引用
基金项目
National Key Research and Development Program of China(2020YFB1707700)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据