| Fabric smoothness degree is an important index for evaluating the fabric performance of appearance and wrinkle resistance.It has an important guiding effect on quality control in the production process and consumer choice in the transaction.Therefore,fabric smoothness degree need to be properly evaluated.At present,the method of sample comparison mainly was used in fabric smoothness evaluation,that is,comparing fabric samples with standard templates at standard illumination through visually compared,and the smoothness degrees are given according to their appearance similarity.The above method is a subjective evaluation method,which is susceptible to the influence of individual psychology,physiology and evaluation environment.The evaluation result has low accuracy and poor stability,and the evaluation result of the same individual may fluctuate.In addition,visual contrast evaluation relies on observers' subjective perception towards smoothness.There is no uniform standard.There may be large differences in evaluation results between individuals,and manual evaluation is time-consuming,labor-intensive,and inefficient,which is not in line with the development trend of intelligent manufacturing in the textile industry.In recent years,the application of machine vision in the field of industrial inspection has been increasing,gradually replacing traditional human eye measurement and judgment,which has brought positive significance to the automatic evaluation of fabric smoothness.Based on this,This paper systematically studies the objective evaluation method of fabric smoothness which mainly includes four parts: the development of image acquisition equipment,frequency domain analysis of fabric images,fabric smoothness feature extraction and pattern recognition model construction.With regard to the development of image acquisition equipment,this paper compares early research and finds that the 3D method has low image acquisition efficiency and high equipment costs,which is not conducive to industrial application promotion.this article determined the research plan for two-dimensional images by comparing early studies,and designed an image acquisition device based on a single-sided illumination environment,and because the wrinkles were randomly distributed on the surface of fabric samples,Sensitive to the illumination environment,the image acquisition device can change the light environment according to needs,which is beneficial to the subsequent discussion on the stability of the illumination environment.Compared with the time domain features of the image,the frequency domain features are more stable,and the information of different frequencies can be studied separately in the frequency domain to reduce the interference of irrelevant information on smoothness feature extraction.In this paper,Fourier transform,Gabor transform and wavelet transform are used to analyze the image in the frequency domain.In the Fourier transform,a gradient filter is set to determine the effective wrinkle contribution interval,and the amplitude information of the wrinkle contribution interval is extracted to represent the fabric smoothness degree.In addition,it was found in the research that low-resolution images can eliminate the effect of fabric structure on the contribution area of wrinkles.In the Gabor transformation,the filter bank is optimized by adjusting the center frequency f,the half-peak bandwidth b,and the filter shape parameter ?,and the value range of the parameter combination(f,b,?)is determined according to the Gabor transformation result.The Gabor filters are sensitive to edges and has a good detection effect on wrinkles.Therefore,a six-direction Gabor filters are set,and its amplitude total(AMT)is calculated to characterize the characteristics of fabric smoothness.In the wavelet transform,it provides a multi-resolution observation and processing tool for the images.This paper performs three-layer wavelet decomposition on the fabric image,observes the wrinkles shape at 4 scales,and obtains 10 images including the original image.A gray level cooccurrence matrix is obtained for each image,and features such as contrast,correlation,and angular second-order moments are used to characterize the fabric smoothness.The results show that the above three characteristics have a highly linear correlation with the level of fabric smoothness and can correctly characterize the degree of fabric smoothness.The objective evaluation of fabric smoothness is a typical pattern recognition,that is,the classification of fabric smoothness is determined based on the extracted feature values.In the pattern recognition model,Support Vector Machine(SVM)is initially selected to classify the smoothness,and an optimized SVM model is obtained through grid search and cross-validation.500 fabric samples(65% training samples and 35% as test samples)was used to verify the effectiveness of this model,the results show that the support vector machine classification accuracy of Fourier features,Gabor features and wavelet-gray co-occurrence matrix features are It reached 78.29%,80.57% and 79.43%,which is a certain improvement over previous research.In order to further improve the stability and accuracy of the model,ensemble learning was applied to the objective evaluation of fabric smoothness.Bagging-SVM model was proposed,and a dual perturbation mechanism was used to improve the diversity of ensemble learning system.Due to the small number of fabric samples,the sample distribution was first changed based on the Bootstrap sampling method to increase the difference between the base learners.Then based on feature selection,the difference between the base learners is further increased.The classification ability of each feature is measured by the information gain,and the feature selection weight is set according to the information gain ratio,and a differentiated feature subset is constructed.The results show that the classification accuracy of the integrated learner for Gabor features and wavelet gray level co-occurrence matrix features is 85.17% and 84.57%,respectively. |
PDF Full Text Request