Fabric Classification Based On Graph Convolutional Network

In the traditional textile industry,the identification of fabric texture was usually based on manual operation,such as sensory method,combustion method,microscope method,dissolution method,dyeing method,etc.These methods not only cannot ensure the accuracy of fabric identification,but also have low efficiency.Some classification and identification methods even need to damage the fabric,which obviously cannot meet the requirements of high automation in the textile industry at the present stage.The combination of computer vision technology and deep learning method is applied to the classification and recognition of fabric material properties,which can greatly improve the production efficiency of the textile industry,and is of great significance to the automatic production of textile enterprises.In this paper,the video data sets of 30 different kinds of fabrics under sustained and stable wind are taken as the research object,and the potential internal relations among different kinds of fabrics are explored by using the swinging multi frame images in the video.Through the fusion of deep learning technology and fabric mechanics model,a new method of describing fabric motion with graph is proposed,The physical properties and mechanical properties of various fabrics in motion state are represented by the characteristics of fabric internal force for recognition and classification.The main work includes:(1)A mechanical model is designed to calculate the interaction force in the fabric,which is called fabric force model.Under the action of wind,the weaving mode and material properties of all kinds of fabrics affect the motion of the fabric.Using the fabric force model with optical flow characteristics of particle advection,the optical flow information of the interaction between particles in the fabric under the motion state is calculated,and the force flow data can effectively extract the force flow characteristics of all kinds of fabric materials in the video.In addition,for the convenience of exploring the potential internal force between all kinds of fabrics The corresponding rules of transformation and selection are put forward to transform the force flow features on the image into digital vector information,which can be used for the later generation of network data of fabric drawing.(2)Because many different kinds of fabrics improve the representation content of fabric network graph,when generating fabric network graph,on the basis of fabric force flow features generated by fabric force model,video time sequence features are integrated.Visual words are made according to the force flow features obtained by fabric force model to explore all potential internal attribute relationships between the same and different kinds of fabrics.In addition,according to the inherent timing characteristics of the video,the fabric nodes in the same video are connected in a certain order to generate the fabric network diagram.(3)According to the characteristics of fabric data set made by fabric force model,this paper proposes a graph convolution neural network architecture for fabric classification and recognition,which integrates multiple types of fabrics and analyzes their force flow characteristics as a whole,and then classifies them.In addition,by adding a graph convolution layer in the graph convolution neural network to obtain more refined and effective fabric feature information,so as to improve the classification accuracy.Tested on MIT's open data sets,the experimental results show that the fabric recognition algorithm proposed in this paper can effectively classify the moving fabrics in the video.Compared with the current mainstream methods,this method is not affected by the surface information such as fabric texture,structure,color and so on.With less memory resources,it can classify the fabric more quickly and efficiently The accuracy of classification is 84.45%.