Research On Fabric Defect Detection Algorithm Based On Deep Convolutional Neural Network

During fabric manufacture,defects will appear on the surface of fabric because of mechanical equipment failure,yarn problems,debris entrapment and human factors,which will bring huge economic losses to textile enterprises.Therefore,fabric defect detection is very important.At present,most textile enterprises in China mainly use two methods to detect fabric defects.One is manual detection,there are low accuracy rate,high rate of missed detection,high rate of false detection,and high labor cost.Second is based on the traditional machine learning algorithm,this kind of method needs to design features manually,but because of the small size,complex shape and large size change of fabric defects,it is hard to extract the features nicely.Convolution neural network has been widely used in industry in recent years.In this paper,deep convolutional neural network is introduced into fabric defect detection field to realize a fast and effective defect detection algorithm,mainly from the following three aspects:(1)Fabric defect recognition algorithm based on deformable convolution is studied.Firstly,feasible data enhancement strategys are come up to solve the sample imbalance between defect classes in data set.Then,due to the fact that many kinds of fabric defects are slender and variable in shape,and the traditional convolution method is not effective for the recognition of such objects.To solve this problem,a fabric defect classification algorithm based on the darknet-53 network is come up,and introduces the deformable convolution to construct the Darknet DCN classification network.The classification accuracy of 96.8% is achieved in the fabric defect test set,and the feasibility of the network is proved by visual analysis.(2)Fabric defect detection based on multi-scale feature extraction network is studied.Two improvements to YOLOV3 are come up.Firstly,in order to reduce the adverse impact of hard to classify samples on the network,the focus loss is innovatively introduced as a loss function to make the model focus on training hard to classify samples.Secondly,it difficult to obtain the feature information of small defects with the deepening of network layers.An improved multi-scale feature fusion network is come up,which combines the high semantic information of deep network with the high resolution of shallow network to obtain more abundant feature information,so as to optimize the defect detection network.When training the model,K-means clustering algorithm is used to initialize the anchor frame.The average detection accuracy of83.6% is achieved on the fabric defect data set,and the model can effectively reduce the miss detection rate and false detection rate of fabric defects..(3)Fabric defect detection based on lightweight network is studied.In response to devices with limited storage or computing capacity,a lightweight fabric defect detection algorithm is come up.To compress and optimize the model of yolov3,the mobilenetv3?large network is used as trunk network to reduce the model parameters.The memory occupied by the model is 101.8MB,the weight of the model is reduced by more than half,and in the test set containing 12 kinds of fabric defects,the m AP reached 80.1%.The practicability of the model is improved.