| In the process of textile production,the inspection of knitted fabric is an important content of quality control while the most important thing is the detection of knitted fabric defects.Because of the modernization of science and technology,the domestic textile production technology has made great progress in the automation application.However,the detection of knitted fabric defects still mainly depends on manual work which has obvious limitations,for example,testers will be easy to fatigue or distraction which may lead to higher rate of error detection and the test result is easily influenced by subjective factors.The automatic detection of knitted fabric defects has restricted the development of textile industry.Nowadays,we should focus on how to use computer vision to detect knitted fabric defects instead of manual vision.In order to make further research on the automatic detection of knitted fabric defects,the paper has studied the the following content :1.Primitive images which were collected by the image acquisition system were changed into grayscale images,the uneven illumination was corrected,images were filtered by median filtering and then changed into binary images.These processes could almost eliminate the texture characteristics of primitive knitted fabric and highlight the defect characteristics.In addition,an innovative method was proposed in this paper which was a rapid online defects detection method based on the gray standard deviation of primitive images and it could improve the efficiency of the test system.2.Adopted db1 wavelet which was compactly supported and orthogonal as the decomposition wavelet and We could get the image of vertical high frequency detail and the image of horizontal high frequency detail after two-time discrete Mallat wavelet transform.The details could clearly reflect the warp texture information and the weft texture information of knitted fabric defects and provide convenience for the later steps.3.When established the gray level co-occurrence matrix,studied the influence of three structural parameters on the four characteristic parameters and chose the best values of structuralparameters through experiments.The IMPROFILE function of the MATLAB was used to get the texture cycle of fabric images.Split the images into panes according to the texture cycle and the structural parameters.Extracted four characteristic parameters from each pane of warp texture images and weft texture images respectively.Because the absolute value of the four characteristic parameters were not in the same order of magnitude and their own fluctuation range were not the same,so the data of all the characteristic parameters should be normalized which could make it easier to train when they were put into the Extreme Learning Machine.4.We selected 50 images of each defect and extracted eigenvalues from each image as experimental data to set up the Extreme Learning Machine training model.After the training of the Extreme Learning Machine,we selected 20 images of each defect to test the accuracy of the training results.The percent of accuracy could be got from the comparison between the test results and the actual results.In order to test the effect of extreme learning machine applied in defect detection and classification,the paper did the following tests :(1)The influence of different number of hidden layer on the classification accuracy of Extreme Learning Machine;(2)the stability analysis of the classification accuracy of Extreme Learning Machine.Through many experiments,the paper determined the number of hidden layer which is the most suitable for this subject and proved that Extreme Learning Machine could maintain a certain degree of stability with the random inputs weights and hidden layer offset.The experiment showed that the method of the defect detection and classification which was proposed in this paper could effectively inspect common defects like holes,pin holes,drop stitch and motion mark in knitted fabric and the accuracy of the identification could reach above 93%. |
PDF Full Text Request