Research On Recognition Of Woven Fabric Structural Parameters Based On Convolutional Neural Networks

Fabric,especially woven fabric,is one of the most important materials in daily life and industrial applications.Its appearance structural parameters mainly include fabric density and weave pattern,and yarn-dyed fabrics also involve the arrangement of colored yarn and color pattern.The efficient and accurate identification of these structural parameters is an important prerequisite for textile enterprises to copy samples and control quality.Nowadays,however,the analysis of fabric structural parameters still needs manual operation.This method is timeconsuming and low efficiency.In recent years,the use of computer vision to recognize fabric structural parameters is increasing,but still no successful system can be applied to practical production.Because the current study is based on traditional methods.The development of the convolutional neural network provides a new thought to the automatic identification of fabric structural parameters.The automatic recognition of fabric structural parameters is systematically studied in this paper.The content of this research main includes:(1)A wireless and portable fabric image acquisition system is developed to solve the problems of unwieldy and complex operation of the existing image acquisition system.The system uses wireless transmission,combined with the RTSP code stream and Web Socket transmission protocol for data exchange.It can be applied to the recognition of fabric structural parameters in commercial scenarios and production environments.(2)A fabric image data set containing more than 400 different types of fabrics is established.The data set contains detailed fabric structural parameters which can be used to train the deep network model,and can also be used as a standard evaluation data set for evaluation.(3)A multi-task and multi-scale convolutional neural network(MTMSnet)is designed for the recognition and analysis of fabric structural features.The network used multi-task structure to locate yarn and floats respectively.At the same time,the multi-scale structure with different convolutional cores is introduced to adapt to the fabric features of different sizes,which improved the adaptability of the network.The fabric density,weave pattern,layout of color yarn arrangement and color pattern are recognized based on the predicted heat maps of fabric structural characteristics.(4)This study designs an online web-based and mobile-based automatic recognition system,respectively.The web-based system uses the Django,Vue.js technology to realize the development in the front and back end.The mobile-based system adopts the Uni-app to realize the cross-platform compatibility.At present,the system has been online for private testing.The results show that: the mean absolute error for fabric density measurement is 1.47%;the weave pattern recognition error is 5.42%;the color pattern recognition error is 10.39%.At the same time,with the help of hot-loading strategy,the computation time is 1.21 s in average.Combined with the image acquisition system and online recognition software system,the detection time is less than 5s.All the results show that the method has strong variety adaptability and reaches the state-of-the-art level,which can be applied in the actual production.