| In recent years,non-woven fabrics are used widely in environmental protection,architecture,and medical domain,and the quantity demanded of automatic equipments are rising in the situation of increasing human cost.A most part of factories which produce non-woven fabrics still on the way of using people’s eyes to check the quality of nonwovens.There is such a large part of factors including people’s subjective judgement and vision fatigue in the process of manual nonwovens inspection that workers always missing the nonwovens defects leading to lower quality of nonwovens.Aiming at the problem of high cost of defect samples acquisition in practical application scenarios,the paper proposes a defect detection system unsupervisedly for nonwovens.The method denoise and extract the effective texture feature with multiple directions.Only by training few normal samples,the model could classify the defect region.Aiming at the high fall-out ratio,the paper improve the OCSVM by adding to local patch weights to balance the level of patches that are more different from the global image and normal patches,improving the accuracy of model and being real-time.Main workload of the paper are as follows:1)The paper explores the way of construct nonwovens defect detection hardware system.According to the application requirements of industrial field and the need of wide field of vision with continuous detection,the paper design a vision system that can achieve the real-time image data acquisition,which construct a type of light path that consist of industrial line-scan digital camera matching industrial highlighted stripe light source by the way of back light.The experiment result shows that the vision system which the paper constructed can meet the actual demand of nonwovens defect detection in the industrial field.2)The paper exerts denoising processing on the image which the visual parts collect and select best methods of extracting features comparing several common ways,which is representing multiple directions’ edge and texture features as support vectors.3)The paper proposes an unsupervised defect detection method based on improved One Class SVM(OCSVM)by designing and optimize OCSVM model aiming at nonwovens defect detection,which maps the support vectors extracted by normal data points from input space into feature space.By training the OCSVM,the model could find a hyper plane to protect the normal points from origin of coordinates and classify the defect regions.The paper proprocesses the training data,simplifies the model,and in view of the situation that the OCSVM after training is easy to misdetect the normal area,the paper modifies the weight of patches whose level are most different of global image,improving the accuracy and real-time performance.4)Based on C#,this paper developed the non-woven fabric inspection software,which includes many camera related parameter modules,detection module and query module,as well as the design of the software interface.The software is easy to operate,and the function can meet the requirements of industrial field use.The experimental results show that the accuracy of the proposed algorithm is more than95.4% in practical application,and the real-time detection speed can reach up to30m/min.Moreover,the algorithm can be used to detect the defects of non-woven fabrics of various specifications. |
