Research On Fabric Defect Detection Based On Deep Convolutional Neural Network

Fabric defect detection is a crucial step in textile production.How to further improve the accuracy and speed of fabric defect detection will have a very positive impact on the production step of textile cloth,which is of great significance to the healthy development of textile industry.Therefore,this paper has carried out a detailed study on the detection accuracy and detection speed in the process of fabric defect detection.The main completed innovative work in this paper is as follows:Aiming at the problem that the state-of-the-art(SOTA)object detection algorithms based on convolutional neural network(CNN)still have room for improvement in the detection accuracy of fabric defects,three effective tricks are proposed in this paper to further greatly improve the detection accuracy of fabric defects.Firstly,the single input image is scaled into several input images of different resolution sizes for multi-scale training,which improves the adaptive ability and recognition ability of the neural network under different scales.Secondly,in order to solve the problem of unbalanced size distribution of fabric defects in fabric data set.In this paper,k-means clustering was carried out on the two dimensions of width and height of defect bounding box in the data set,and then the fixed prior boxes in the object detection algorithm were replaced by the clustering centers,which makes the learning process of network model easier.Finally,by replacing the traditional non-maximum suppression(NMS)method with the improved non-maximum suppression method(Soft-NMS),this paper effectively eliminates the phenomenon of repeated detection of the same type of defects in the data set.Experimental results show that the accuracy of fabric defect detection based on baseline model is improved by 8.9% m AP(mean Average Precision)through continuous stacking of these tricks.Aiming at the problem of repeated detection and possible misjudgment in the ways of detection speed and effect in the state-of-the-art object detection algorithms,a novel defect detection network(Defect Net)is proposed to solve the problem of defect detection.Firstly,this paper compares the difference between the detection of defect data and the detection of general natural object data.It is pointed out that there are a large number of images without annotations in the defective data,and its processing principle is fundamentally different from the general target detection problem,which makes it clear that the application of the general object detection algorithm based on convolutional neural network may not be perfect in this problem.Then,the defect network method proposed in this paper is verified and analyzed by experiments.It uses the defect finding network to quickly determine whether an image contains defects by sharing convolutional operation with the backbone network in the object detection network model,and then decides whether to conduct defect detection on the image according to the judgment results.If the image does not contain defects,the final detection result without defects is directly returned;otherwise,further defect detection is required.Theoretical derivation and experimental results fully demonstrate the detection efficiency and effects of the Defect Net method proposed in this paper.Taking the data set in this paper as an example,the proposed Defect Net method effectively improves the detection speed of object detection network by 29.2% and F1-score by 7.2% m AP without affecting the detection accuracy of the original network model.