Research On Fabric Defect Detection Method Based On Convolutional Neural Network

Fabric defect detection is an important link of production and quality management in textile industry,the current manual testing has some shortcomings,such as slow speed,high labor intensity,subjective factors and lack of consistency.Traditional machine learning algorithms for fabric defects are usually used to deal with features that are easy to extract and quantify,For example: color,area,roundness,angle,length,etc.however,the defect features can not be extracted well because of the small proportion of defect areas,large variation of defect scale and complex defect shape.Convolutional neural network is a new and important method in the field of image recognition.In this paper,convolutional neural network is applied to fabric defect detection and recognition to improve the speed and accuracy of recognition.Firstly,the characteristics of fabric defect data set are analyzed.According to the corresponding characteristics,the corresponding data enhancement method is adopted,the composition of residual network is analyzed,the first generation ResNet50 model is established,and the baseline is obtained by experiment.Secondly,Focal Loss and multi-task assisted Loss are adopted to solve the problem of unbalanced fabric defect categories.Aiming at the small proportion of defect areas,a single-scale patch block sampling strategy is proposed.The composition of SE module is introduced.A SE-ResNet50 model based on SE module is established,and the corresponding test methods are proposed to improve the detection accuracy..Then,aiming at the problem of large scale change of defect area,the characteristics of defect area scale change are analyzed,a multi-scale patch block sampling strategy is proposed,a SE-ResNet50 model is established,and the corresponding test method is proposed,which improves the detection accuracy.However,due to the use of a large number of patch blocks,many regions have repeated feature extraction,which greatly reduces the detection speed.Using the improvement of Fast-RCNN relative to RCNN,the convolution feature is extracted only once from the original image,and the convolution layer feature map is pooled at multi-scale.A multi-scale pooling model(MPP Net)and corresponding test method are proposed.The experiment shows that the model can effectively solve the problem of too slow detection speed without affecting the detection accuracy.Finally,the work done in this paper is summarized,and several directions and application fields that can be continued in the future are given.