Research On Image Classification Of Fabric Defects Based On Dictionary Learning

The image classification method of fabric defects is a research hotspot in the field of intelligent textile industry.In recent years,classification methods based on sparse dictionary learning have received extensive attention and have been applied to the task of fabric defect image classification.However,traditional sparse dictionary learning based classification methods have the following three problems:1)The feature information extracted from the training samples is not rich and the discrimination capability of the extracted features is not strong;2)Sparse problem solving is time-consuming,complex and inefficient;3)The sparsity and discrimination of the representation coefficient are insufficient.These problems lead to the poor performance of classification methods based on sparse dictionary learning in fabric defect image classification.According to the above problems,this thesis carries out related exploration and research,and proposes an improved sparse dictionary learning method,and a simple classifier can be used in the classification stage to achieve better classification performance.The research content and related work of this thesis are summarized as follows:（1）A fabric defect image classification method based on fast category constraint dictionary learning（FCCDL）is proposed.Firstly,On the traditional sparse dictionary learning model,this method constructs a dictionary learning model for fabric discriminant features by suppressing intra class differences and inter class fuzziness.Then,L₂regularization is applied to the coding coefficient instead of L₁regularization in the traditional dictionary learning method,which effectively reduces the computational complexity and improves the classification efficiency.Finally,through the sparse representation of the test samples,the reconstructed error vector is obtained for classification.In addition,the experimental results on TIANCHI fabric defect data set and the fabric data set collected by the factory in real time prove the effectiveness of the fabric defect image method.（2）A fast classification constraint dictionary learning method based on linear discriminant analysis（LDA-FCCDL）is proposed for fabric defect image classification.In terms of the advantages of the FCCDL model,linear discriminant constraints are applied to the representation coefficients to make the distance between the sparse coefficients of the same class decomposition as close as possible and the distance between the sparse coefficients of different class decomposition as far as possible in the dictionary learning process.The sparse representation model is reconstructed to obtain a more discriminative dictionary.At the same time,samples of the same category can be represented by specific dictionaries,but not by specific dictionaries of different categories,so as to minimize intra class differences and increase inter class differences.The dynamic interpretation is that the subtle differences of the same category encourage the finding of representative dictionaries in this category,and the large differences of different categories inhibit the sparse representation of samples of other classes by the representative dictionaries,and better sparse representation performance is obtained.In addition,the experimental results on TIANCHI fabric defect data set and the fabric data set collected by the factory in real time prove that compared with the existing sparse dictionary learning classification method,this fabric defect image classification method has stronger discrimination and better classification performance for fabric defect images.（3）A fabric defect classification system based on LDA-FCCDL is designed,which makes it easier for users to use the fabric defect image classification method,and further verifies the effectiveness of the fabric defect image classification method based on LDA-FCCDL.