基于改进3E-LDA的织物图像分类算法

摘要/Abstract

摘要： 针对训练样本数太少（训练样本数量小于数据维数）导致的模型分辨能力下降问题，提出了一种基于正则化改进3E-LDA的织物图像分类算法，称为I3E-LDA算法。首先利用类加权中值代替样本均值计算类内散点矩阵，削弱了离群值和噪声的影响。以此作为非参数加权特征提取法对类内散点矩阵进行正则化。然后利用目标组合的方法，通过引入平衡参数对目标函数进行正则化，来保留更具判别性的特征数据。通过不同织物图像间更具判别性的特征数据可以更好地对其区分。结合改进的零空间法解决类内散点矩阵奇异性和小样本问题，从而提高分类准确率。在阿里天池织物数据集和花色织物图像上进行训练和测试，将图像按照正常图像和非正常图形（瑕疵图像）进行区分。实验结果表明，I3E-LDA算法有效实现了织物图像分类，且对于较少的训练样本（20%~40%的样本用于训练）提升了分类精度。

关键词: 线性判别分析, 织物, 图像分类, 正则化, 小样本

Abstract: Textiles are one of the three physical elements of clothing. In recent years, with the development of computer technology, the classification and recognition of fabric images and intelligent manufacturing have played a very important role in the textile field. In the process of production, traditional manual detection methods are still widely used for fabric defect detection, which is time-consuming and laborious, and the efficiency is very low. It is easy to cause false detection and missed detection due to fatigue, and it will also affect the quality and price of textiles. Therefore, the use of digital image processing technology to complete the recognition and classification of fabric images has become a hot issue in recent years.
Relying on machine vision, spots, pits, scratches, color differences and defects on the fabric surface can be detected. Linear discriminant analysis (LDA) is a supervised dimensionality reduction and classification algorithm that can effectively classify fabric images. LDA classifies by calculating the optimal discriminant matrix to minimize the intra-class distance and maximize the inter-class distance. However, LDA is sensitive to outliers, ignores local geometric information and small sample size (SSS), which affects the classification accuracy. The 3E-LDA (three enhancements to linear discriminant analysis) algorithm improves the above three problems on the basis of LDA and improves the classification accuracy. However, when the number of training samples is smaller than the data dimension, it will reduce the model's resolution ability and ultimately affect the classification accuracy.
A fabric image classification algorithm based on improved 3E-LDA, called I3E-LDA algorithm (Improved 3E-LDA), was proposed to address the problem of reduced model resolution caused by training samples being smaller than the data dimension. Firstly, the nonparametric weighted feature extraction (NWFE) method was used to regularize the intra-class scatter matrix, and then the goal combination method was used to introduce equilibrium parameters to regularize the objective function, so as to weaken the influence of outliers and noise, retain more discriminative feature data, and rely on these feature data to better classify fabric images. It is necessary to combine the improved null space learning method to solve the singularity and small sample problems of intra-class scatter matrices and improve classification efficiency, and to train and test on the Alibaba Tianchi fabric dataset and pattern fabric images to distinguish between normal and abnormal patterns (defect images). The experimental results show that the I3E-LDA algorithm effectively achieves fabric image classification, and improves classification accuracy for a small number of training samples (20%-40% of the samples are used for training).

Key words: , linear discriminant analysis, fabric, image classification, regularization, small sample size

中图分类号:

TP181

靳文哲, 吕文涛, 郭庆, 徐羽贞, 余润泽. 基于改进3E-LDA的织物图像分类算法[J]. 现代纺织技术, 2024, 32(6): 89-96.

JIN Wenzhe, LÜ Wentao, GUO Qing, XU Yuzhen, YU Runze. Fabric image classification algorithm based on improved 3E-LDA[J]. Advanced Textile Technology, 2024, 32(6): 89-96.

参考文献

[1]吕文涛, 林琪琪, 钟佳莹, 等. 面向织物疵点检测的图像处理技术研究进展[J]. 纺织学报, 2021, 42(11): 197-206.
LÜ Wentao, LIN Qiqi, ZHONG Jiaying. Research progress of image processing technology for fabric defect detection[J]. Journal of Textile Research, 2021, 42(11): 197-206.
[2]田宸玮, 王雪纯, 杨嘉能, 等. 织物瑕疵检测方法研究进展[J]. 计算机工程与应用, 2020, 56(12): 8-18.
TIAN Chenwei, WANG Xuechun, YANG Jianeng, et al. Research progress on fabric defect detection methods[J]. Computer Engineering and Applications, 2020, 56(12): 8-18.
[3]YIN W, MA Z. High order discriminant analysis based on Riemannian optimization[J]. Knowledge-Based Systems, 2020, 195: 105630.
[4]LU R, CAI Y, ZHU J, et al. Dimension reduction of multimodal data by auto-weighted local discriminant analysis[J]. Neurocomputing, 2021, 461: 27-40.
[5]HU X, SUN Y, GAO J, et al. Probabilistic linear discriminant analysis based on L1-norm and its Bayesian variational inference[J]. IEEE Transactions on Cybernetics, 2020, 52(3): 1616-1627.
[6]CHANG W, NIE F, WANG Z, et al. Self-weighted learning framework for adaptive locality discriminant analysis[J]. Pattern Recognition, 2022, 129: 108778.
[7]YAN C, CHANG X, LUO M, et al. Self-weighted robust LDA for multiclass classification with edge classes[J]. ACM Transactions on Intelligent Systems and Technology, 2020, 12(1): 1-19.
[8]LI C N, QI Y F, ZHAO D, et al. F F-norm two-dimensional linear discriminant analysis and its application on face recognition[J]. International Journal of Intelligent Systems, 2022, 37(11): 8327-8347.
[9]高云龙, 王志豪, 丁柳, 等. 动态加权非参数判别分析[J]. 控制与决策, 2020, 35(8): 1866-1872.
GAO Yunlong,WANG Zhihao, DING Liu, et al. Dynamic weighted nonparametric discriminant analysis[J].Control and Decision, 2020, 35(8): 1866-1872.
[10]陆荣秀, 蔡莹杰, 朱建勇, 等. 基于自权值线性判别分析算法的图像处理研究[J]. 华东交通大学学报, 2020, 37(3): 135-142.
LU Rongxiu, CAI Yingjie, ZHU Jianyong,et al. Application of self-weight linear discriminant analysis in dimension reduction[J]. Journal of East China Jiaotong University, 2020, 37(03): 135-142.
[11]HUANG Libo, LING Yongquan. A parameter-free local linear discriminant analysis method[J]. Computer Science and Application, 2021, 11: 1042.
[12]NIE F, ZHAO X, WANG R, et al. Fast locality discriminant analysis with adaptive manifold embedding[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(12): 9315-9330.
[13]梁志贞, 张磊. 面向 Kullback-Leibler散度不确定集的正则化线性判别分析[J]. 自动化学报, 2022, 48(4): 1033-1047.
LIANG Zhizhen, ZHANG Lei. Regularized linear discriminant analysis based on uncertainty sets from kullback-leibler divergence[J]. Acta Automatica Sinica, 2022, 48(4): 1033-1047.
[14]YU Y F, REN C X, JIANG M, et al. Sparse approximation to discriminant projection learning and application to image classification[J]. Pattern Recognition, 2019, 96: 106963.
[15]LI Y, LIU B, YU Y, et al. 3E-LDA: Three enhancements to linear discriminant analysis[J]. ACM Transactions on Knowledge Discovery from Data , 2021, 15(4): 1-20.
[16]GAO J, LI L. A robust geometric mean-based subspace discriminant analysis feature extraction approach for image set classification[J]. Optik, 2019, 199: 163368.
[17]ZHENG N, GUO X, TIE Y, et al. Incremental generalized multiple maximum scatter difference with applications to feature extraction[J]. Journal of Visual Communication and Image Representation, 2018, 55: 67-79.