Research on fabric classification based on graph neural network

• Published in: Industria textilă (Bucharest, Romania : 1994) Vol. 74; no. 1; pp. 3 - 11
• Main Authors: TAO, PENG, WENLI, CAO, JIA, CHEN, XINGHANG, LV, ZILI, ZHANG, JUNPING, LIU, XINRONG, HU
• Format: Journal Article
• Language: English
• Published: Bucharest The National Research & Development Institute for Textiles and Leather - INCDTP 01.01.2023
• Subjects: Agglomeration; Classification; Fabrics; Feature extraction; Graph neural networks; Graphical representations; Neural networks; Teaching methods; Textile industry; Textiles; Wavelet transforms
• ISSN: 1222-5347
• DOI: 10.35530/IT.074.01.202224

Fabric classification plays a crucial role in the modern textile industry and fashion market. In the early stage, traditional neural network methods were adopted to identify fabrics with the drawback of restricted fabric type and poor accuracy. Combining multi-frame temporality and analysing fabric graph data made from fabric motion features, this paper proposes a novel hybrid model that introduces the concept of graph networks to classify 30 textile materials in a public database. We utilize the graph inductive representation learning method (GraphSAGE, Graph Sample and Aggregate) to extract node embedding features of the fabric. Moreover, bidirectional gated recurrent unit and layer attention mechanism (BiGRU-attention) are employed in the last layer of aggregation to calculate the score of previous cells. Intending to further enhance performance, we link the jump connection with adaptive selection aggregation frameworks to determine the influential region of each node. Our method breaks through the limitation that the original methods can only classify a few fabrics with great classification results.