基于Transformer网络特征融合的色纺织物图像检索
Image retrieval of colored spun fabrics based on Transformer network feature fusion
沈佳忱 1袁理 2廖海斌 1王闵 1郭旻1
作者信息
- 1. 武汉纺织大学 电子与电气工程学院,湖北 武汉 430200
- 2. 武汉纺织大学 电子与电气工程学院,湖北 武汉 430200;武汉纺织大学 省部共建纺织新材料与先进加工技术国家重点实验室,湖北 武汉 430200
- 折叠
摘要
针对单一特征难以准确描述色纺织物图像颜色和纹理的复杂性和各向异性,提出了基于Transformer网络特征融合的色纺织物图像检索算法.首先,利用卷积神经网络模型提取色纺织物图像的高级语义信息;然后,利用Transformer网络融合图像的三阶颜色矩特征,充分利用色纺织物图像的"高级语义信息"和"浅层图像信息"的互补性进行图像检索.采用了 14 种不同类型的色纺织物样本图像进行检索,该系统的平均Top-10 查全率与准确率mAP值分别达到了 98.35%和 89.25%,相较于融合注意力机制的单一网络模型,检索Top-10 的查全率和mAP值均有提升.
Abstract
Aiming at the complexity and anisotropy of the color and texture of colored spun fabrics images that are difficult to be accurately described by a single feature,a color spun fabrics image retrieval algorithm based on Transformer network feature fusion was proposed.First,the high-level semantic information of colored spun fabrics was extracted by using the convolutional neural network model;then,the third-order color moment features of the image wewere fused by using the Transformer network to make full use of the complementarity between the ＂ high-level semantic information＂ and the ＂ low-level semantic information＂ of colored spun fabrics for image retrieval.Then,the third-order color moment features were fused using the Transformer network to make full use of the complementary nature of＂high-level semantic information＂ and ＂low-level semantic information＂ of color textile images for image retrieval.In this paper,14 different types of color-spun knitted fabric sample images were used for retrieval,and the Top-10 recall and mAP of this system reach 98.35%and 89.25%,respectively,which is an improvement in the recall and mAP of retrieving Top-10 compared to a single network model incorporating an attention mechanism.
关键词
色纺织物/图像检索/Transformer网络/特征融合Key words
colored spun fabrics/image retrieval/Transformer network/feature fusion引用本文复制引用
出版年
2024
NETL
NSTL
万方数据