The Defect Detection Algorithm Of Denim Fabric Based On Transfer Learning

In China,the production and total export volume of denim clothing have always ranked first in the world,but the competitiveness of China's denim clothing industry shows obvious weakness.One of the main reasons is that the defects in the production process of denim fabric have not been detected.Manual visual inspection is a common method for traditional denim fabric defect detection,which is inefficient and prone to omissions and false detection.Detection methods based on vision and image processing require high image acquisition,low detection accuracy,slow processing speed,and poor robustness,so they have not been applied to actual industrial production.Aiming at this problem,in this paper,transfer learning is adopted to study the defect detection algorithm of denim fabric to realize the online detection of defects.The work of this paper is carried out mainly in the following aspects:(1)Based on the analysis of domestic and foreign research status of fabric defect detection,the idea of denim defect detection based on transfer learning is formed.In addition,based on the analysis of the structural characteristics of the denim fabric,as well as combined with the denim fabric defect types to statistically analyze the denim fabric image data set,eight types of denim fabric defects to be detected and the distribution rules of the defects are sorted out.(2)Combined with the characteristics of denim fabric defects,a preprocessing method for denim fabric defect data sets is proposed,including enhancement of image data,expansion of sample data sets,and based on this,the diversity of denim fabric defects is improved.(3)On the basis of transfer learning,a dual model fusion detection algorithm for denim fabric defects is proposed.On a large data set ImageNet,the algorithm first extracts the weight parameters of the pre-trained VGG16 model,and trains the fabric defect detection classifier and the defect recognition classifier using its portability.Then on the selected denim fabric defect data set,the weight parameters of some convolution layers of the two models are retrained and adjusted,and finally the models are fused.Through comparison simulation experiments,the results show that the dual model detection algorithm is compared with the VGG16 model,ResNet50 model,and Xception model.After comparison simulation experiments,the results show that compared with the VGG16 model,the ResNet50 model and the Xception model,the dual-model detection algorithm achieves an accuracy rate of 94.3% in detecting 8 kinds of common denim fabric defects such as knots,three threads,rough wefts,holes,and warp warps,and the algorithm is more robust.(5)In this paper,a prototype of the denim fabric defect detection system is set up,and the key parts of the prototype,such as the light source,camera,and image processor,are introduced.In the prototype,the detection algorithms based on the dual-model fusion detection algorithm based on transfer learning,the detection algorithms based on the VGG16 model,the ResNet50 model,and the Xception model are compared.According to the actual measurement of the prototype,it can be concluded that the dual-model transfer detection algorithm shows high detection accuracy and detection speed in detecting 8 kinds of common defects in denim fabric.Based on transfer learning,the defect detection algorithm of denim fabric is researched.Combining the structural characteristics of denim fabrics with common defect types,a pre-processing method of defect image sample data and a dual model fusion detection algorithm are proposed,and the rationality and superiority of the dual model fusion detection algorithm are verified through comparative simulation and prototype testing respectively.