Fabric Texture Stability Representation And Its Applications Based On Dictionary Learning

For fabric structure parameters and defects,texture has a direct influence on their automatic detection and identification.The effective characterization of the fabric texture is of great significance to the promotion of the core competitiveness of traditional textile enterprises.At present,the fabric texture analysis is still completed by manually for most of the textile companies in China.However,such identification and inspection relying entirely on human experience and vision can no longer meet the requirements of modernization and automation.Compared with manual analysis,texture analysis based on computer vision technology not only can efficiently improve the detection accuracy,but also can objectively evaluate the appearance and inherent quality of textiles,and is consistent with the development trend of automation and intelligence in the textile industry.The previous research has illustrated that dictionary learning method can efficiently describe fabric texture,but the obtained model is not unique.Moreover,the texture stability representation is the fundamental study of texture analysis and application.So,this paper presents the definition of the fabric texture representation focusing on the investigation of the texture stability characterization algorithm and the corresponding evaluation indexes.The specific research contents and relevant conclusions are as follows:(1)A new algorithm for global texture stability representation based on dictionary learning is designed.According to literature studies,it is believed that the initial dictionary is fixed and it should be able to obtain a stable representation of the texture.Taking into account the characteristics of the woven fabric itself,this paper takes the DCT dictionary as the initial one and adopts an alternating least squares approach to characterize fabric texture images by dictionary learning.That is,the least square method is used to get the coefficient matrix and update the dictionary,alternately update the coefficient matrix and dictionary,then the learned dictionary can be obtained after several iterations.The stability of the proposed algorithm is analyzed and illustrated,and verified by experiments.The experimental results of the nine global original images show that there is no difference between the reconstructed images and original ones by naked eye,and the average values of mean reconstruction quality and root mean square error of the sample set are 31.46 dB and 8.08 respectively.The analyzation of image size impact on texture representation illustrates that under the same parameters,the smaller the image size,the better the texture characterization result.Therefore,the non-overlapping sub-window partition is introduced,and dictionary learning is performed on the sub-window sample image.The fabric sample images are represented based on the global and local patch size separately.It can be concluded that the global way can represent the texture than the small local patch,namely the reconstructed quality increases 12 dB,and root mean square error reduce 11.This is because the whole image is divided into small patches by non-overlapping,which ignores the correlation between the sub-windows sample images,and greatly weakens texture features such as texture directionality when characterizing fabric images.In addition,the global and local-based noisy images are also characterized.The results show that these two ways have poor performance in noisy images texture representation and are not suitable for texture representation of noisy images.(2)A local fabric texture representation algorithm using fixed dictionary is presented.Since the global algorithm is limited to the original image,no effective texture characterization can be performed for noisy images,and the redundant information of the image is not fully utilized,the non-overlapping division method cannot reconstruct the overall texture features well.So,a local image-based texture characterization algorithm is designed with the non-overlapped partitioning method.The proposed algorithm uses the DCT dictionary as the fixed dictionary,and then the coefficient matrix for small patch samples through OMP algorithm.The reconstructed image can be derived from the DCT dictionary and coefficient matrix.After experiment analysis optimizes dictionary size and sparsity cardinality,and select the appropriate small patch size for original and noisy images,meanwhile proposes a TXC index for fast optimized patch size for noisy image.For nine sample sets and their noisy images,the experimental results indication that the average reconstruction quality and the structural similarity of the original images is 36.90 dB and 0.98 respectively.In comparison with the global algorithm,the proposed method has better representation result.For noisy image,the moderate noise research results show that when TXC > 0.5,the small patch size 16? 16 has better characterization result than 8? 8,otherwise,the small patch size 8? 8 is better.The designed method can effectively represent? =15,20 noisy image with appropriate patch size.Compared with NLM approach in? =10,the proposed has higher reconstruction quality.(3)Aiming at the fixed dictionary poor adaptability to fabric texture,the concept of the general dictionary is proposed,and the algorithm of local fabric texture representation based on the general dictionary is designed.The purpose of the general dictionary is to improve the universality and adaptability of the dictionary.Due to the efficiency of K-SVD dictionary learning,this paper adopts it to train the selected sample images to get the general dictionary.In this article,training three kinds of general dictionarires.The first one is an ordinary general dictionary,which the training samples is freely choose.The second type is organizational structure general dictionary,and there are four this kind of dictionary based on different weave patterns such as twill,weft satin,honeycomb pattern and diamond twill.The third one is a joint organizational structure general dictionary(diamond twill weave,the weft satin,basket and honeycomb).Nine sample data sets results display that under the same conditions,compared with the fixed dictionary,general dictionary for the reconstruction of the original image quality raises 0.25 ~ 1.63 dB,structural similarity increases 0.002 ~ 0.03.Their noisy image test result shows that general dictionary for part of the reconstruction quality of noisy image increases 0.33 ~ 1.30 dB and increases the structural similarity of 0.001 ~ 0.02.The rest of the samples of reconstruction quality is lower via the proposed method than the fixed dictionary 0.03 ~ 0.21 dB,and the similarity is reduced by 0.001.It can be seen that the general dictionary has better adaptability to fabric texture,and it is better to represent the original image than the noisy image.On this basis,the general dictionary,four weave repeats general dictionaries and joint organization structure general dictionary represent diamond twill,the weft satin,basket and honeycomb images,and the their characterization results from good to bad in the order: organizational structure general dictionary > joint organization structure general dictionary > ordinary general dictionary.This means that the specificity of the organizational structure dictionary and the joint organizational structure dictionary is better,and the ordinary general dictionary has a better general universality.(4)The influence of fabric structure parameters on the texture representation using general dictionary is investigated.Mainly investigation the impact of weave density and weave repeat on texture representation.For 32 fabric samples,experimental results show that the reconstruction quality general became better as the weaving density increased within the scope of fabric density.8 different weave patterns(plain,twill,warp satin,weft satin,basket,honeycomb,complicated twill,diamond twill)experimental results indicate that the diamond twill,honeycomb and complicated twill weave repeat are easily affected by the fabric density.That is to say,with the change of the fabric density,the characterization results produced large fluctuations.Within the density scope in this paper,the fabric density has a great influence on fabric texture,every weave pattern has its own applicable density range.Within this range,the texture features of the fabric structure are more uniform and the characterization is better,and the differences in various organizational structures will be clear.(5)This paper presents a flaw detection algorithm based on general dictionary.Due to the area size and of the defect is different and diverse flaw types,in this paper,the size of the sub-window is optimized,and then the defective samples are tested.According to the different sub-window sizes 16? 16,26? 26,32? 32 and 36? 36,the K-SVD dictionary learning method is used to train the normal fabric texture to obtain the different general dictionaries.Then the defect fabric sample is divided into small patch in none-lapping way,moreover patch samples are converted into a column vector,every patch is equal to a column vector.Afterwards,the patch defect sample to be tested is reconstructed by the ordinarily general dictionary and the residual matrix is computed.In the end,the defect sample is detected column by column via thresholding method,and the defect region was marked on the black rectangle as the final detection result.The optimization result of the four patch sizes shows that the sub-window 26? 26 has the lowest false detection rate for linear defects,non-linear defects,and mixed defects,next to the sub-window 16? 16 with the best detection rate.Given that the difference between the detection rates of the two is less than 1%,and the false detection rate of the sub-window 26? 26 is the lowest,hence the selection of the 26? 26 quantum window size is a good compromise.The experimental results of plain and twill weave pattern defect samples(49 in total,17 for each of linear and non-linear defect samples,and 15 for mixed defect samples)show that the algorithm has a good adaptability to fabric texture and defect type,and obtain 95.2% detection rate and 2.5% false detection rate.For the other complex organizational structure defects(6 in total,3 for liner defects and 6 nonlinear defects),the detection rate is 97.2 and the error detection rate was 0.In addition,the training dictionary is trained done in the offline state,which does not affect the speed of the whole detection.Research conclusion of this paper,provides reference and theoretical basis for the storage of fabric images and the establishment of its digital model,and offers a feasible scheme for the basic analysis of fabric texture,and supports a reference for realizing on-line fabric inspection.