Research On Non-woven Fabric Defect Classification And Detection Method Based On Deep Learning

With the increasing social demand for non-woven fabric,non-woven fabric has become an emerging sunrise industry in the 21 st century and is widely used in the fields of health care,clothing,architecture,packaging and automobile.However,in the production process of non-woven fabrics,due to environmental and human factors,non-woven fabrics are prone to quality problems such as dirty spots,folds,yarn loss and breakage,These problems affect the production quality and efficiency of enterprises.The traditional manual detection method has the problems of low efficiency,high error detection rate,high price of professional equipment detection method and difficulty in customized detection.In view of the above problems,deep learning technology is applied to non-woven image defect classification detection,so that the accuracy,precision and average detection time of an image sample can meet the basic requirements of industrial real-time detection.The research object of this paper is non-woven fabric sample pictures.The classification detection model is obtained by deep learning method,and the model is applied to the classification detection of non-woven fabric defect images.Firstly,the data sets needed for deep learning are constructed.Secondly,three classical convolutional neural networks are selected and the model is trained.The deficiencies of the model are found through experiments.Finally,an improved convolutional neural network model is proposed.The model meets the requirements of enterprises in the accuracy,precision and detection time of classification detection.The main research contents are as follows:(1)Select the classic GoogLeNet,RestNet and MobileNet,and adjust the above three networks appropriately.Two training methods were adopted: the ab initio training model with sufficient samples and the transfer learning training model with a small amount of samples.The experimental results show that the two training methods have little effect on the results of non-woven fabric defect classification.The accuracy and precision are only up to 95% of the detection requirements of ResNet from scratch training and MobileNet from transfer learning training.The detection time of each of the above three networks fails to meet the requirements of real-time detection.(2)In order to reduce the detection time of the convolutional neural network model,AlexNet network with small model depth was selected.In the process of AlexNet network training non-woven fabric image,an improved method of sample image preprocessing was proposed to solve the problem of large model computation and difficult convergence.The methods of image geometric transformation,image median removal and normalization were used to preprocess the sample image.The experimental results show that the method of image de-median and normalization can significantly improve the classification detection accuracy and precision of AlexNet network.(3)In order to further improve the classification and detection effect of AlexNet model,a structural improvement method is proposed.On the basis of the original model,increase the number of convolution layers.After the adjustment of various parameters and the comparative analysis of a large number of experimental data,a model suitable for the classification and detection of non-woven sample image was obtained.The accuracy and precision of the improved model are greatly improved in the classification and detection of non-woven fabric defects.The experimental results show that the accuracy and accuracy of the improved model are over 95% and the detection time is within 35 ms,which conforms the requirements of the factory for the non-woven fabric defect classification detection.