A Study On Fabric Defect Detection Methods Based On Intelligent Learning Algorithm

Detection of fabric defects has long been performed manually.The low detection rate,which is caused by poor working conditions,high labor intensity,high false and missed detection rate,and experience-based detection criteria,is unable to meet the demands of modern textile industry.With the development of computer technology,digital image processing technology and machine vision technology in recent years,it has become a trend to replace the tradition manual detection with machine vision detection.Machine vision detection of fabric defects includes image acquisition,pre-processing,image segmentation and defect features extracting,classifier design,detection and classification.Certain achievements are gained through study on these topics,and defect detection and classification methods are designed for woven fabric based on intelligent learning algorithms.This paper mainly focuses on the following aspects.(1)Image pre-processing.The original color images are converted into gray scale images at first step.Due to the low contrast between the defect area and the environment background of the gray scale images,fabric images are enhanced through gray level adjustment.To distinguish the defect area,an ideal lowpass filtering method in the frequency domain is designed.The experiment results show that this method is of strong anti-disturbance capacity and filters successfully that the fabric background information can be effectively removed and the defect area is then distinguished.(2)Classification design of woven fabric defects based on BP(Back Propagation)neural network.Roberts operator is used for edge detection and segmentation of the filtered gray-enhanced image.A BP neural network classifier based on multiple features is designed.The experiment shows that most defects are effectively classified.(3)Classification design of woven fabric defects based on deep learning.To improve the accuracy of identification and classification,deep learning is studied.A new MDML deep learning network model for fabric defect detection and classification is constructed using multi-dimensional filters and multi-pooled techniques based on the VGG(Visual Geometry Group)convolutional network model.The experiment shows that MDML network model achieves high recognition efficiency and accuracy,which improves the detection and classification performance of deep learning network for woven fabric defects.(4)Adaptive boosting of BP neural network and deep learning network.An adaptive boosting algorithm for weak classifiers is constructed to obtain a strong classifier.In light of the imbalance in the samples of conventional adaptive boosting algorithm and classifier performance reduction,a new design method is suggested using dynamically regulated training sample set and introducing the improved Sigmoid function to restrain sample weight.As a result,a new adaptive boosting algorithm AdaBoost.WSA for multiclass classification is designed and realized.The experiment shows that AdaBoost.WSA algorithm effectively solves the above mentioned problems and improves the sampling balance and classification efficiency.(5)Design of experiment system and construction of classifier performance evaluation method.An experiment system for woven fabric defect detection and classification using intelligent learning algorithms is designed and commonly encountered woven fabric defects are studied in the experiment.Criteria of classifier performance evaluation method which is suitable for woven fabric defect detection and classification is constructed.Data analysis,evaluation method and calculation formulas are illustrated in detail.Data information is collected through dynamic defect detection on the experiment platform,and the evaluation method is applied to the comparative experiment of several algorithms.The experiments show that this evaluation method is suitable for the performance evaluating of woven fabric defect detection and the modified algorithm improves both the accuracy and efficiency of classification.