Research On Fabric Defect Detection Algorithm Based On Fast Lightweight Convolutional Neural Network

Fabric defect detection is an important link in the process of textile production,which plays an important role in improving the quality of textile production.The traditional fabric defect detection is mainly completed by the inspector,which has the problems of slow speed and high rate of missing detection,and can not meet the requirements of fast and high quality production.In order to improve the efficiency of fabric defect detection,it is necessary to use automatic detection technology instead of tedious manual detection to improve production efficiency and fabric quality.Due to the diversity and complexity of defects in fabric texture image,automatic detection of fabric defects is a challenging task in the field of machine vision.In recent years,convolutional neural network(CNN)has developed rapidly in the field of target detection,and its detection ability for various types of targets has been generally recognized.CNN algorithm has been applied to the detection of fabric defects.However,with the improvement of the detection accuracy of the CNN model,the calculation cost and storage requirements of the model are also significantly increased,which to some extent hinders the application of the CNN model in the environment with limited computing and storage resources,such as mobile or embedded devices.Therefore,this paper studies the algorithm of fabric defect detection based on CNN,and optimizes the model structure according to the texture characteristics of fabric defect.The model can improve the accuracy of fabric defect detection,at the same time,reduce the model parameters and improve the detection speed.The innovative research results of this paper are as follows:1)A fabric defect detection model based on LW-YOLO lightweight convolutional neural network is proposed.Firstly,the feature map from each layer of the neural network is visualized,and the depth of the network model is determined by analyzing the feature information extracted from each layer.A light-weight convolution neural network framework is constructed by continuous 3?3 and 1?1 convolution layers.Then,the feature pyramidnetwork is used to realize multi-scale feature extraction,so that the model can extract different fine-grained feature information and improve the detection ability.Finally,the coordinate value of each defect is calculated accurately by using the logistic regression algorithm to locate the defect target.The experimental results show that compared with the YOLOv3 model,the model can compress the model without affecting the detection accuracy,and the parameters of the model are significantly reduced.2)The DefectNet convolutional neural network is proposed to detect fabric defects at high speed.Firstly,the convolution process is divided into two parts: depthwise convolution and point convolution.The feature extraction and feature fusion of the feature map are realized respectively,which reduces the calculation amount of the convolution process in theory,significantly reduces the calculation amount of the convolution operation,and improves the detection speed.Then,K-means clustering algorithm is used to get the initial anchor frame size of fabric defects,which improves the ability of defect location of the model.3)A new fast feature extraction block called "separable adaptive recalibration module(SAR)" is proposed,which combines depthwise separable convolution and attention mechanism,and introduces a new "feature recalibration" strategy.The importance of each feature channel can be acquired automatically by learning.The network can use global information to selectively enhance the useful feature channel and suppress the useless feature channel,so as to realize the adaptive calibration of feature channel and improve the ability of network feature extraction.