Intelligent characterization and evaluation of yarn surface appearance using saliency map analysis, wavelet transform and fuzzy ARTMAP neural network
Introduction
The evaluation of yarn surface appearance, or so-called yarn grading, is an important routine in assessing yarn quality in textile industry. The appearance of yarn not only significantly affects its commercial value but also directly influences the characteristics of end products (Grover & Hamby, 1966). Generally, the appearance quality of yarns highly depends on the regularity of yarn surface. A greater unevenness in yarn surface implies poorer quality. The evaluable defective items or yarn faults may include thin place, thick place, nep, fuzz and foreign matter. The first three refer to the irregularities in yarn diameter: a short yarn segment with thickness apparently less than or greater than the average yarn diameter. Fuzz represents those fiber ends that protrude from the yarn surface. A less number of such fiber ends indicates a smoother yarn. Foreign matter mainly refers to a small number of fibers or trash of different materials to the majority of others.
In order to evaluate the yarn surface appearance, a standard test method has been established by ASTM D2255 (ASTM D 2255-02, 2007). In this standard method, a yarn specimen is continuously wound onto a rectangular or trapezoidal-shaped black board. The board covered with the wound yarn is then examined and a visual appraisal of appearance is made. Traditionally, the inspection is carried out by direct observation in which a skilled specialist visually compares the wound table with photographic standards labeled in Grades A, B, C and D, and then assigns a grade to the yarn sample. The photographic standard of Grade A represents the best grade quality and the others are progressively lower. The general descriptions for each grade can also be provided in terms of yarn fuzziness, neps, unevenness, and visible foreign matter (ASTM D 2255-02, 2007):
- •
Grade A: good uniformity, good cover, no large nep, no excessive fuzziness.
- •
Grade B: a few small neps, small pieces of foreign matters, a bit fuzzes and irregulars.
- •
Grade C: more neps, more fuzziness, greater contrast between the thick and thin places.
- •
Grade D: large neps, more fuzzes and more foreign matters, over-all rougher appearance.
This method has been used successfully for years to grade textile yarns. However the manual method of inspection is subjectively dependent on human vision, thus it is much skill-oriented, judgmental and inconsistent. Therefore, it is desirable to develop an objective, accurate, and automatic method for the evaluation of yarn appearance.
Recently, computer vision and artificial intelligence have attracted the increasing attention of researchers in textiles and apparel. Attempts have been made to replace the conventional observation method with computer technology to resolve the limitation of human vision. The notable applications include fiber morphological measurement (Bel and Xu, 2010, Huang and Xu, 2002, Wan et al., 2009, Wang et al., 2010, Wang, Yang et al., 2010, Wong et al., 2009, Xu and Huang, 2004), yarn snarl characterization (Xu et al., 2008, Xu et al., 2010), fabric pilling evaluation (Kim and Kang, 2005, Palmer and Wang, 2003), fabric texture and defect recognition (Kuo and Tsai, 2006, Lin, 2009, Murrells et al., 2009, Shady et al., 2006, Xu and Bel, 2009, Yuen et al., 2009), and nonwoven uniformity identification (Liu, Zuo, Zeng, Vroman, & Rabenasolo, 2010). For instance, Liu et al. introduced a Learning Vector Quantization (LVQ) neural network to evaluate nonwoven uniformity (Liu et al., 2010). Liu’s model has achieved an average accuracy of 87.7% in the identification of five nonwoven grades from 625 nonwoven images. To classify garment defects in the apparel industry, Yuen et al. proposed a novel hybrid model with a combination of Genetic Algorithm (GA) and a three-layer Back-Propagation (BP) neural network. In this model, GA was used to optimize the shape of fabric structuring element and BP was employed for image classification process (Yuen et al., 2009). GA algorithm was also employed by Hsu et al. to schedule yarn-dyed production process in textile, which has been shown to be superior to the most commonly used EDD scheduling method (Hsu, Hsiung, Chen, & Wu, 2009).
In evaluation of yarn surface appearance, Rong and Slater (1995) developed a microcomputer system for yarn unevenness analysis. In this system, the analogue signal of the diameters of a single yarn, measured by a standard Uster Tester, was converted into digital form and then further analyzed in terms of statistical parameters of unevenness. Similarly, Neva, Lawson, Gordon, Kendall, and Bonneau (1994) and Nevel, Avser, and Rosales (1996) proposed an electronic system to read the diameters of a single yarn when it moves over a CCD camera. The whole length of yarn was then split into a number of shorter lengths so that they can be displayed electronically side-by-side to assist in manual grading. In the system, some yarn defects, such as thick places and neps, can also be identified and counted, and the number of defects along the length of yarn is used as a grading index for evaluating yarn surface appearance. In all the above-mentioned methods, although it was possible to define a classification for yarn surface appearance, a grading method based on a standard image of yarn board is found to be impossible. Most importantly, they are all based on the mechanism similar to the conventional Uster Tester and depart from the originality of the visual grading method defined in ASTM D2255, namely the comparison among adjacent yarn segments and the identification of other important yarn surface appearance features, such as foreign matters and hairiness, being ignored.
More recently, Semnani et al. introduced an inspection method by using image processing and artificial neural network (Semnani, Latifi, Tehran, Pourdeyhimi, & Merati, 2006). In this method, a simple threshold method was repeatedly used to separate thick places from yarn body and then the background from the thick places. Although the method somehow enables an evaluation from a standard yarn image, the yarn inclination on the black board will have a great influence on the accuracy of the method. In addition, the local thin places could not be recognized because the feature has to be removed from thick places together with yarn body in the algorithm. Neither foreign matters nor hairiness were identified in this method because of its inherent limitations. Therefore, there is a need to solve these challenges and develop a new accurate and robust intelligent system for the objective evaluation of yarn surface quality.
In the present paper, an intelligent characterization and evaluation model of yarn surface appearance is presented. The mechanisms of manual inspection of yarn board are studied and an integrated intelligent method for yarn feature extraction and classification is then proposed. Some image processing and pattern recognition approaches, including saliency map, wavelet transform and fuzzy ARTMAP neural network, are employed and incorporated for the simulation of manual inspection and the objective evaluation of yarn surface appearance. The rest of this paper is organized as follows. Section 2 depicts the methodology of our proposed model, including saliency map, wavelet transform and fuzzy ARTMAP neural network. Section 3 shows the experimental results on the yarn image database. Finally, the concluding remarks are given in Section 4.
Section snippets
Methodology
The proposed objective grading system (see Fig. 1) is mainly composed of three modules, namely image pre-processing, yarn characterization module and fuzzy ARTMAP neural network. First, an image pre-processing method is utilized to binarize yarn board images for further processing. Secondly, a yarn characterization model is proposed to extract yarn characteristics from binary yarn line images and yarn hairiness images. In this module, the yarn board image is analyzed by using three image
Image database
In our experiments, 576 yarn board images with five categories of yarn count (20 Ne, 32 Ne, 40 Ne, 60 Ne and 80 Ne) are used. Yarn count is a special term used in the textile for specifying the linear density or thickness of yarns. The larger the value of yarn count, the finer the yarn thickness. Therefore, in this yarn image database, thickness of normal yarn lines is varied for different linear densities, where 20 Ne is the thickest yarn lines and 80 Ne is the thinnest yarn lines. The resolution of
Conclusion
In this paper, an intelligent grading system is proposed for textile yarn evaluation and grading based on image processing technique and the fuzzy ARTMAP neural network. An attention-driven fault detection method is proposed to detect visual difference among yarn segments. Wavelet transform technique is utilized for extracting hairiness characteristics under a certain decomposition level using four types of mother wavelets. Furthermore, other important yarn features, such as yarn diameter
Acknowledgement
The authors would like to thank the Hong Kong Polytechnic University for funding supports of this work (Project No: 1-BB9U and A-PD1F).
References (60)
- et al.
A signal denoising algorithm based on over complete wavelet representations and Gaussian models
Signal Processing
(2007) - et al.
Attention-driven image interpretation with application to image retreival
Pattern Recognition
(2006) - et al.
An efficient algorithm for attention-driven image interpretation from segments
Pattern Recognition
(2009) - et al.
A GA methodology for the scheduling of yarn-dyed textile production
Expert Systems with Applications
(2009) - et al.
A fuzzy neural network approach to machine condition monitoring
Computers and Industrial Engineering
(2003) Neural network based detection of local textile defects
Pattern Recognition
(2003)Applying wavelet transform, rough set theory and support vector machine for copper clad laminate defects classification
Expert Systems with Applications
(2009)- et al.
Nonwoven uniformity identification using wavelet texture analysis and LVQ neural network
Expert Systems with Applications
(2010) - et al.
Finding out general tendencies in speckle noise reduction in ultrasound image
Expert Systems with Applications
(2009) - et al.
Incipient gear box fault diagnosis using discrete wavelet transform (DWT) for feature extraction and classification using artificial neural network (ANN)
Expert Systems with Applications
(2010)
A new wavelet-based image denoising using undecimated discrete wavelet transform and least squares support vector machine
Expert Systems with Applications
Stitching defect detection and classification using wavelet transform and BP neural network
Expert Systems with Applications
Fault gear identification and classification using discrete wavelet transform and adaptive neuro-fuzzy inference
Expert Systems with Applications
Application of a modified fuzzy ARTMAP with feature-weighted learning for the fault diagnosis of bearing
Expert Systems with Applications
A novel fault diagnosis method of bearing based on improved fuzzy ARTMAP and modified distance discriminant technique
Expert Systems with Applications
A hybrid model using genetic algorithm and neural network for classifying garment defects
Expert Systems with Application
Fiber longitudinal measurements for predicting white speck contents of dyed cotton fabrics
Textile Research Journal
Texture analysis and classification with tree-structured wavelet transform
IEEE Transaction on Image Processing
Genetic evolution processing of data structures for image classification
IEEE Transactions on Knowledge and Data Engineering
An efficient learning algorithm for adaptive processing of data structure
Neural Processing Letters
A fuzzy ARTMAP fault classifier for impulse testing of power transformers
IEEE Transactions on Dielectrics and Electrical Insulation
A decision support system for wool classification
AUTEX Research Journal
Overcomplete steerable pyramid filters and rotation invariance
Proceedings of IEEE Computer Vision and Pattern Recognition
Image analysis for cotton fibers: Part I: Longitudinal measurements
Textile Research Journal
Cited by (38)
Functionalized nanofibers for piezoelectric energy harvesting applications
2023, Functionalized Nanofibers: Synthesis and Industrial ApplicationsMachine learning and soft computing applications in textile and clothing supply chain: Bibliometric and network analyses to delineate future research agenda
2022, Expert Systems with ApplicationsCitation Excerpt :Cluster 6, shown in purple colour, revolves around the theme of “Pattern Recognition and Clustering of Yarns and Fabrics”. These studies mostly involve assessment of visual characteristics of yarns and fabrics, or retrieval of coloured patterns from fabric with the help of techniques like BP-ANN, ANFIS, GA, convolutional neural network (C-ANN), support vector machine (SVM) and fuzzy C-means clustering (Eldessouki & Hassan, 2015; Liang et al., 2012; Xiao et al., 2018). Cluster 7, coloured in brown, incorporates articles related to “Colour Prediction and Matching”.
Yarn Tensile Properties Modeling Using Artificial Intelligence
2020, Alexandria Engineering JournalCitation Excerpt :Image processing was used to obtain yarn board images. saliency map, wavelet analysis, and statistical measurement were used to extract yarn faults and Hairiness, and a fuzzy neural network was trained using the acquired features [11]. Sheng Yan Li et al utilized a two scale attention system and probabilistic neural networks for yarn surface grading.
Hierarchically architected polydopamine modified BaTiO<inf>3</inf>@P(VDF-TrFE) nanocomposite fiber mats for flexible piezoelectric nanogenerators and self-powered sensors
2020, Nano EnergyCitation Excerpt :Piezoelectric nanogenerators (PENGs) and triboelectric nanogenerators (TENGs) exhibit excellent capability to efficiently convert different forms of mechanical energy to electric energy, including human body motion [15–17], vehicle movements [8,18,19], water flow [20–22], air flow [20,23,24], and acoustic wave etc. [25,26] Compared with TENGs, PENGs have the advantages of lower dimension, compact structure, high sensitivity and durability, thus have attracted considerable attention as energy harvesters, sensor networks and self-powered systems [1]. However, with the development of flexible electronics in wearable devices, implantable biomedical devices and smart textiles [27–31], the commonly used piezoelectric inorganic semiconductors (e.g., ZnO, GaN and InN), and piezoelectric ceramics (e.g., BaTiO3(BT), ZnSnO3, PbZrxTi1-xO3, (K, Na)NbO3 cannot meet the demand of adequate flexibility due to their intrinsic frangibility [32]. Because of their merits of excellent flexibility, easy processing and lightweight, piezoelectric polymers such as polyvinylidene fluoride (PVDF) and their copolymers [e.g., poly (vinylidenefluoride-co-trifluoroethylene), P(VDF-TrFE)] show promising potential in flexible and wearable applications [33–35].