Research On Fabric Defect Detection Method Based On Convolution Neural Network

The textile industry has always been one of my China’s traditional pillar industries.Currently,in order to ensure fabric quality,one of the key issues in the textile industry is regarding how to conduct high-quality detection of fabric defects.Aiming at the problems of poor feature extraction capability,complex network models,and missed or error detection of small target defects such as pilling,holes,and stains in the current fabric defect detection algorithms,this paper focuses on the study of fabric defect detection algorithm based on convolutional neural networks.The main research content includes the following aspects:(1)Firstly,aiming at the problem of poor feature extraction capability of the current fabric defect detection algorithm based on convolutional neural network,this paper proposed an improved Faster R-CNN feature extraction network algorithm based on six kinds of defective fabric datasets collected in the laboratory.According to the characteristics of three main defects,coarse yarn,broken yarn and tension line,which account for a large proportion of the datasets in this paper,the Resnet50 feature extraction network in the original Faster R-CNN was improved to Resnext-101-64 to reduce the complexity of the network model and improve the model’s ability to extract deep features in this paper.Experimental results show that the m AP of this model after improving Resnext-101-64 reaches 0.879,which is greatly improved compared with the original Faster R-CNN.However,the missed detection rate of defects such as pilling,holes and stains is still high.(2)Secondly,in order to reduce the missed detection rate of pilling,holes and stains in the model,through visual result analysis,it was found that the corresponding defects were all small target defects.So an improved simple two-way fusion PANet was further proposed to extract more information and improve the breadth of the network,thereby enhancing the detection capabilities of small target defects.Experimental results show that the detection accuracy of the model in this paper has been improved again.Especially for the above three small target defects,the missed detection rate of has been greatly reduced.Meanwhile,the detection accuracy of the other three types of defects has also been improved to a certain extent.(3)Finally,in order to further improve the average detection accuracy of the model in this paper,the CIOU LOSS regression loss function is adopted after analyzing the characteristics of several IOU loss functions.After a series of ablation mixing experiments,it has been demonstrated the improved mode’s superior performance in terms of m AP and average detection accuracy of six types of defects either compared with the original two-stage target detection models Faster R-CNN and Cade R-CNN or the single-stage target detection models SSD and Yolo X.This verifies the effectiveness of the improved model in this paper on fabric defect detection problems.