Research On Fabric Classification Based On Graph Neural Network

In the present day of information technology,the rapid change of textile industry and people’s increasing consumer demand drive the fabric classification recognition technology to keep pace with the times.The traditional fabric detection technology has the limitations of single type of textile material and low recognition accuracy,while graph neural network gradually becomes a hot spot in various research fields.Therefore,it is considered to find a breakthrough to update the means of fabric classification from the level of graph structure and explore a low-cost,efficient and accurate fabric recognition and classification method.In this thesis,the research is carried out on a real fabric dataset of thirty species.For fabrics in motion,fabric feature information is obtained based on the fabric mechanics model,while a fabric graph network is constructed by combining video temporality.Based on this,a research method based on hybrid graph neural network is proposed.The main work is as follows.(1)For the fabric feature extraction problem,independently of the attributes of fabric surface color,texture and tissue composition,the force flow characteristics of the moving fabric are obtained from the physical mechanics perspective through the fabric force model.At the same time,the potential internal relationships among various types of fabrics are explored by combining the inherent multi-frame timing characteristics of the video.The traditional Euclidean fabric video is converted into non-Euclidean fabric graph data to form a graph network structure carrying fabric objects and their relationship information.(2)Combining the data characteristics of the fabric graph,the graph inductive learning representation method(Graph SAGE)is used for the classification and recognition of fabrics,and the rich embedding representation information of itself is obtained by sampling and aggregation operations to neighboring fabric nodes.For the aggregator,which plays a key role in the graph inductive learning representation method,there is a limitation of fixed neighbor radius in the aggregation process,which cannot achieve the optimal vector representation of fabric nodes.Therefore,a jump connection mechanism is introduced to achieve the goal of adaptive adjustment of the aggregation radius of fabric nodes.It enables both sparse nodes located at the edge of the fabric graph and dense nodes at the center of the graph network to have the ability to aggregate the rich embedding information of their neighbors.Combined with the bi-directional gating cycle unit to refine the key node information,the importance score of each aggregation layer for a specific fabric node is calculated by the layer attention mechanism,and the final effect of various fabric node clustering is obtained.The hybrid graph neural network model proposed in this thesis is validated on a publicly available real fabric dataset,and the results show that this method significantly reduces the recognition time cost,while the accuracy rate reaches 95.9%.