Research On Defect Detection Of Texture Fabric Based On Convolutional Neural Network

With the development of the current society,people's living standards keeping improve,and the requirements for clothing are getting higher and higher,which has prompted the textile industry to continuously carry out technological innovation and technological progress.Defect detection is absolutely necessary for the realization of product character assessment in the textile industry production process.At present,China mainly adopts manual methods for the detection of defects in fabric.This method not only has a high false detection rate,a high cost,and a slow detection speed,but also incompatible with the people-oriented concept,the rapid and accurate detection of defects in fabrics through the combination of automation technology and computer vision technology is of great significance to the upgrade of the entire industrial chain.Based on computer vision technology,this thesis conducts research on the detection of defects in textured fabric,and aims at solving the problems of the wide variety of defects in textured fabric and the large differences in shape and size distribution,and the great interference of texture background in detection.While studying the high-precision fabric defect detection algorithm,the detection speed of the model is optimized to realize the automation of cloth defect detection.The main work and results of this thesis are as follows:(1)Analyze the fabric defect data set,filter the samples in the data set,and remove the samples that cannot be used for training.Aiming at the problem of insufficient data in the data set,the data amount is expanded by means of paste enhancement and mosaic enhancement,while the imbalance of various types of defects in the data set is alleviated to a certain extent.At the same time,the traditional image processing method is used to suppress the complex texture in the fabric and eliminate the interference of the complex texture.(2)Construct a fabric defect detection model based on the YOLOv5 m network.For the problem of the category imbalance in the data set and the different difficulty of detection,optimize the design on the loss function.Refer to the idea of Focal Loss to deal with the difficulty in the loss function.Easily divided samples and samples of each category are given different weights to balance the weight of various samples in the model training process.At the same time,DIOU is used as the evaluation standard to improve the weighted non-maximum suppression,which reduces the missed detection rate of the model to some extent.(3)In order to achieve better detection results,a two-stage Faster R-CNN network is used to construct a fabric defect detection model.Adjust and optimize the residual module in Res Net50,and then replace VGG16 in Faster R-CNN as Backbone to improve feature extraction capabilities.The feature pyramid network is added to the network structure to integrate features of different scales to improve the model's ability to detect different defects.At the same time,it is proposed to use additive fusion and convolution fusion in the feature pyramid network to reduce the redundancy in the texture fabric.(4)There are various types of textures in cloth,and the interference of different types of textures on defect detection is not the same.Here,the influence of different textures in the defect detection process is studied.Experiments have shown that complex textures will indeed cover the model's extraction of defect features.At the same time,the generalization performance of the model on different texture cloths is verified.(5)Although the two-stage model can get higher precision,its parameter amount and forward speed are not as good as the single-stage model.This thesis adopts the idea of compact network design to optimize the depth separable convolution,and then apply it to the proposed network,which can effectively reduce the amount of model parameters and calculations.At the same time,in order to further speed up the model inference,the method in Thi Net is used to prune the convolution kernel in the network.The parameter amount and inference time of the final model are greatly reduced,which basically meets the real-time detection requirements of fabric defects.