Research On Defect Detection Algorithm Of Patterned Fabric Based On Deep Learning

At present,the mainstream quality inspection method for the fabric in factories is manual inspection in the textile field.With the development of industrial intelligence,machine vision has gained more attention in the textile field with its advantages of fast detection speed and high precision.Some textile enterprises begin to use the defect detection system based on machine vision,and the current defect detection system is mainly used for the defect detection of pure color cloth.According to the actual production of fabric in the factories,combined with the consumer demand for clothing cloth,it found that most of them is the patterned fabric with complex background pattern.Because the defect feature extraction of patterned fabrics is interfered by complex background patterns,the traditional machine vision algorithm relies on the features of manual design to extract features,so it cannot achieve the good effect in the defect detection of patterned fabrics.Therefore,it is of great significance to study the algorithm used in the defects detection of patterned fabrics.First,observe the production process and manual inspection of the patterned fabric in the factory.Seven common defect types of patterned fabrics are selected,the characteristics of patterned fabrics are summarized,and the difficulties in defect detection of patterned fabrics are analyzed.Based on the feature that background pattern has great influence on defect feature extraction in color and other aspects,color linear array camera is selected and its matching lens,light source and other components are selected to build an online detection platform for fabric defect.By comparing the advantages and disadvantages of the two mainstream models of deep learning target detection and combining with the actual requirements of fabric defect detection,a two-stage model suitable for the current detection situation is selected to help complete the design of the detection model infrastructure.Then,the data set of fabric defects needed for deep learning training is established.Collect images of fabric defect obtained from factories,write image clipping programs,and use the data set labeling tool to establish a data set for fabric defect.The public data set of fabric defects is used,and the program is written to convert the data set into the format of data set built by myself,so as to expand the sample size of fabric with defect types needed in the actual detection process.Then,in order to improve the accuracy of the defect detection model of patterned fabrics,the geometric transformation data enhancement method is used to expand the defect data set of patterned fabric.The model effect is enhanced through the preprocessing step of the detection algorithm,and the network training effect is improved through the steps of image preprocessing,migration learning of the backbone network,and the optimization of the anchor parameters in the two-stage detection model by using k-means clustering method.An algorithm suitable for detecting patterned fabric defects is designed.The detection model is improved and optimized by adding deformable convolution optimization backbone network,design of multi-scale model,construction of cascade network structure,introduction of online hard sample mining algorithm and optimization of loss function.Finally,the offline detection experiment of fabric defect images and online detection experiment of the patterned fabrics are carried out.In the offline detection experiment,the fabric defect data set is divided into training set and test set,and multiple sets of fabric defect detection experiments are set up.Through the statistical analysis of the data in the model training process,the visual display of the defect detection effect,the statistical analysis of the test results that the average accuracy is 94.62%,the validity of the algorithm in this paper is verified.On the online test platform of the patterned fabrics set up in the laboratory,the online test of the patterned fabrics is carried out.Experimental results show that the algorithm has good detection effect.The average accuracy rate is about 93% and the detection speed is 0.6m/s on the experiment of patterned fabrics.It can meet the needs of the factory for the defect detection of patterned fabrics.