Research Of Fabric Defect Detection Based On Deep Learning

Fabric defect detection plays an important role in textile quality control,detecting fabric defects in time and controlling fabric production from the source can effectively reduce the losses of enterprises.The fabric defect detection method based on deep learning not only ensures the accuracy,but also the detection velocity is quick,so it has become the primary choice of textile enterprises.However,due to the fabric image has the characteristics of complex structure and texture and various types of defects,the convolution neural network is easy to ignore the characteristic information of defects in the convolution layer,this results in problems such as poor network detection rate and poor classification.Aiming at the above problems,this paper mainly completes the following research work:(1)A fabric defect detection method based on improved Refine Det is proposed.Aiming at the poor classification performance and inaccurate positioning results of the traditional Refine Det model,it is improved.Firstly,VGG16 is changed into full convolution network to extract the features of fabric image;Secondly,in order to obtain the important features of defects and suppress unnecessary features,an attention mechanism is added to the Anchor Refinement Module(ARM);To enhance the classification capabilities of the network,SE module(Sequence and Exception,SE)is added to the Transfer Connection Block(TCB);Finally,the Object Detection Module(ODM)returns the detected results to the correct target position,then predicts the category of defects and locates them.The experimental results show that the average accuracy map of the four categories of fabric images of hole,stain,yarn faults and thread is 79.7%,which is 5% higher than that of the traditional Refine Det detection method,and has good classification and positioning effect.(2)A fabric defect detection algorithm based on improved DCGAN(Deep Convolution Generative Adversarial Networks)is proposed.Firstly,aiming at the problem that the generation layer in DCGAN network only adopts one layer of deconvolution,which makes the network unable to recognize the small defects in the fabric image,the global perception module and attention mechanism are innovatively introduced into the original DCGAN to improve the original DCGAN.In this method,the global perception module is added to the generation layer of the original DCGAN,so that the network can fully obtain the global information of the fabric image and avoid losing the detail information of the defect;Secondly,in order to make the generator pay more attention to defects and give higher weight to the defect area than other areas,the attention module is introduced into the generator,which helps the model better identify the characteristic information of different types of defects;Finally,the target detection module returns the detection results to the correct target position,predicts the defect category and locates the defect.The experimental results show that the algorithm achieves 95.52%accuracy in classifying seven different types of fabric images,which is 3.17% higher than that of the original DCGAN.