基于图卷积神经网络的织物分类研究

织物的分类研究在织物生产、服装设计等领域有着广泛应用。提出织物力模型,结合多帧时序信息和图卷积神经网络,给出一种用图来描述织物运动的新方法,通过分析和提取织物视频中的运动特征,实现织物的分类。该方法使用30种不同织物在风力吹动下的视频作为实验数据集,将视频每一帧作为一个图节点,然后根据视频时序性连接同类织物节点的边。此外结合织物力模型对原视频图像作预处理以提取力流特征作为视觉单词存储,再依据视觉单词探索同类与不同类织物间的潜在联系,由此将欧氏织物视频数据转换为非欧氏织物图数据,最后使用图卷积神经网络进行分类处理。该方法避免了传统织物分类过程中织物纹理、颜色、外部光照等因素的影响,突破了传统分类方法只能对少数织物进行分类的限制,有较好的分类效果。

Fabric classification based on graph convolutional network

Fabric classification research has been widely used in the field of fabric production,clothing design and so on.This paper introduced a novel method combining fabric force model,the multi-frame timing information with GCN for fabric classification.This method used 30 different fabrics’moving videos with the wind driving as the experimental datasets.Firstly,it took each frame of the video as a graph node,and then according to the video sequence connectioned for the similar fabric edge nodes.Secondly,it used the fabric force model to preprocess the original video image to extract force flow features as visual words and store them,then used the visual words to explore the potential connections between similar and different types of fabrics,so as to transform the Euclidean fabric video data into non-Euclidean fabric graph structured data.Finally,these data processed by GCN network.This work has a good performance without the effectiveness of light,texture and color,breaks through the limitation that traditional classification method can only classify a few fabrics,and has good classification effect.