| At present,the traditional texture analysis of woven fabrics mainly depends on artificial vision,which has the problems of strong subjectivity and low efficiency.With the gradual transformation of textile industry to intelligent direction,the application of image processing technology and machine vision to analyze fabric texture features has become a research hotspot.In recent years,dictionary learning has been introduced into the sparse description of woven fabric texture.Research showed that dictionary learning could effectively represent texture features but the existing dictionary learning models were only suitable for the reconstruction of single texture corresponding to each type and there was little research on the impact of texture classification on dictionary reconstruction results.Therefore,based on the sparse representation theory under the comprehensive model,this paper constructed a sparse representation model of discriminant shared dictionary for a large sample size,focused on comparing the effectiveness of two clustering algorithms based on sparse representation theory in woven fabric texture segmentation and applied them to the sparse representation model of discriminative-shared dictionary.The specific research work of this paper is as follows:（1）This paper constructed a sparse representation model of woven fabric texture based on structured discriminative-shared dictionary.The model was divided and sampled intensively in the form of sub-window and selected plain,twill,satin,plain variable weave,twill variable weave and other woven fabrics of two warp and weft yarn densities as the research samples.The research results showed that the structured discriminative-shared dictionary in this paper could characterize the texture of various types of woven fabrics with multiple samples and the effect was good.（2）Based on the above texture representation model,the number of discriminative subdictionaries and shared-dictionaries required to reconstruct the texture image were optimized by taking plain,twill,satin,plain variable weave,twill variable weave and other woven fabrics with different warp and weft yarn densities as test samples.By comparing the differences of peak signal-to-noise ratio,structural similarity index measure and corresponding discriminative indexes of all reconstructed textures in ten groups of classification schemes,it was proved that warp and weft yarn density and weave structure would affect the reconstruction performance of sparse representation model,and the latter had a more significant impact on the recognition ability of discriminative sub-dictionary.（3）According to the sparse representation idea of structured discriminative dictionary learning,6)-way spectral clustering was selected as the division method of woven fabric samples and the clustering results were quantitatively evaluated based on the internal validity evaluation index of clustering.The results showed that the clustering effect of 6)-way spectral clustering algorithm was only slightly improved compared with K-means clustering,but the clustering scheme was not suitable for the sparse representation of structured discriminative-shared dictionary model.（4）This paper proposed a clustering method based on SPCAAPK-means.It was found that SPCAAPK-means performed better than other algorithms when selecting distance metrics to evaluate the clustering effectiveness.Based on the clustering results of SPACAPK-means,the sparse representation model of discriminative-shared dictionary was applied to characterize the woven fabric texture image.The results showed that the reconstructed image contained almost all the texture features of the original sample image and the recognition ability of the total structured-dictionary was good.The reconstruction quality of samples with the same weave structure and warp and weft yarn density was significantly better than other classification results.（5）Based on SPCAAPK-means clustering,the difference of reconstruction effect obtained by sparse representation model of shared dictionary was distinguished after fixing warp and weft yarn density label and weave structure label respectively.The results showed that the texture representation ability of discriminative sub-dictionaries was better than the reconstruction effect obtained by direct clustering,regardless of the fixed density label or the fixed weave structure label.Compared with the fixed density label,the recognition ability of the obtained dictionary would be higher and more stable.The conclusions of this paper can guide the texture characterization,classification and reconstruction of woven fabrics. |
