Classifying fabric defects with evolving Inception v3 by improved L2,1-norm regularized extreme learning machine
Zhiyu Zhou,Xingfan Yang,Jiangfei Ji,Yaming Wang,Zefei Zhu
DOI: https://doi.org/10.1177/00405175221114633
IF: 2.3
2022-07-28
Textile Research Journal
Abstract:Textile Research Journal, Ahead of Print. To improve efficiency and classification accuracy, and overcome the issue of poor generalization performance of traditional fabric defect classification methods, we present a new fabric defect classification algorithm with evolving Inception v3 by improved L2,1-norm regularized extreme learning machine. Herein, first, the features of fabric images are extracted using Inception v3, which reduces the amount of computation and improves the accuracy of feature extraction. Second, an L2,1-norm regularization extreme learning machine algorithm based on multi-verse optimizer–Henry gas solubility optimization, called MHL21-RELM, is proposed. The proposed algorithm solves the problems of the low classification accuracy of the traditional ELM and prevents the algorithm from easily falling into the local optimal solution. Next, instead of SoftMax, MHL21-RELM is used as the classification network layer of Inception V3, and a new fabric defect classification algorithm named Inv3-MHL21-RELM is proposed to enhance the fabric defect classification performance. Finally, the performance of our algorithm is verified by experiments on different fabric defect datasets. The proposed method achieves 97.29% classification accuracy with the Textile Texture Database; it achieves a 6–30% improvement in classification accuracy compared with other traditional fabric defect classification methods, thus effectively improving the classification accuracy of fabric defects. In addition, this method achieves good classification results with different types of fabric defect datasets, thereby showing strong generalization.
materials science, textiles