Research On Fabric Defect Detection Teachnology Based On Deep Learning

Our country is a major producer and consumer of textiles,and there are many textile enterprises.Affected by the external environment and the needs of its own development,enterprises are required to improve product quality to ensure their competitiveness.In all processes,the detection is the last checkpoint of guaranteed product quality,so it needs to meet the requirements for accuracy and efficiency at the same time.The efficiency of manual detection of cloth defects is low and the detection quality cannot be guaranteed;the cloths detected by traditional image algorithms are mostly plain cloths,which are not applicable to a wide range;therefore,in order to detect defects in patterned cloths,this thesis selects the volume of deep learning.As a research method,the cumulative neural network aims to design a set of algorithms to detect the defects of patterned fabrics.The research contents are as follows:(1)The network structure has been improved accordingly.In view of the material characteristics of the cloth and the large size of the entire cloth,this thesis selected the two-stage model as the experimental model.After that,Faster RCNN was selected for training and testing.The network parameters are initialized by a pre-training model,which is trained by other data set with large amounts.For the purpose of obtaining a better detection effect,this thesis selects Cascade RCNN instead of Faster RCNN for detection;considering the interference of cloth color on detection,after the experiment,the method of subtracting the feature map of normal cloth and defective cloth is adopted to suppress the color to improve detection.Besides,in order to further improve the detection accuracy,this thesis improves the loss function and sampling method used by Cascade RCNN.Compared the result of Faster RCNN,this training result improves mAP by 6%.(2)Analyze the data set.Expand the data set for the problem that the number of some types of defects in the data set is not enough.the data is augmented by various method,including translation,rotation,random crop.Besides,in order to enhance the detection effect of small defects,the improved Copy-Paste method is applied to increase the number of defects.Compared the result of no use data augument when training Cascade RCNN,this training result improves mAP by 3%.(3)In view of the problem that the detection effect of small-sized defects and different shapes defects is unsatisfactory,this thesis uses an attention mechanism in improved Cascade RCNN to process the information in the feature pyramid.regarding the feature pyramid as a three-dimensional tensor,using the attention mechanism for the three dimensions respectively,and finally combining three dimensions of attention mechanism to process the feature pyramids.Finally,the Acc of the improved model for cloth defect detection achieved 92.1%,and the mAP achieved 57.2%.Compared with the existing algorithms,the algorithm in this thesis improved the detection accuracy of various defects,and also significantly improved the detection effect of categories with many small-sized defects such as contamination and flower hair.