Abstract:

Objective Fabric pilling grading is a crucial aspect of fabric quality inspection. With the continuous increase in production levels within the textile industry, fabric pilling detection still relies on manual methods, which struggle to keep pace with production efficiency and are susceptible to subjective influences. Consequently, a portable fabric pilling grading evaluation system has been developed to address the demands of rapidly growing production efficiency for both efficiency and accuracy in the inspection process.
Method In this study, a detachable specimen clamp was designed and light angles were optimized, which collectively culminated in the development of a portable fabric acquisition device. The grading evaluation system employed an improved DA-Unet network model based on U-Net architecture to perform semantic segmentation on pilling areas of both knitted and woven fabrics made from cotton and polyester materials. Through correlation analysis, the characteristic parameters of fabric pilling are determined, and the final pilling grade evaluation module is constructed by using machine learning classification algorithm and numerical analysis of related characteristic parameters.
Results The system uses the DA-Unet network model to semantically segment the fabric image under different incident light angles, and compares the obtained semantic segmentation results with the results of pilling areas marked by professionals. Through the four evaluation indexes of intersection over union, pixel accuracy, recall and precision, the optimal incident light angle is determined to be 60°, and the results of the four evaluation indexes are 75.83%, 98.75%, 79.59% and 94.67%. The influence of characteristic parameters including pilling number, total pilling area, maximum pilling area, average pilling area, median pilling area and contrast, on pilling grade of fabrics was explored. The correlation between the characteristic parameters and the grade was analyzed by Spearman and Kendall methods, and the four characteristic parameters that had the greatest influence on the grade evaluation were determined as pilling number, total pilling area, maximum pilling area and median pilling area. Because of the differences in pilling characteristics between knitted and woven fabrics, six machine learning classification algorithms are used to predict the grade of four characteristic parameters of the two fabrics (combined with the grade evaluation results as input data). The results show that the random forest classification algorithm has the highest accuracy on knitted and woven fabrics, reaching 97.42% and 95.31% respectively. Because the special circumstances such as the appearance of hairballs in a large area have a great influence on the evaluation results, the related numerical range of different grade characteristic parameters is further analyzed, and the grade evaluation module is finally completed through the different grades divided by the parameter values. When analyzing the grades of 1 200 knitted and woven fabric samples, with discrepancies maintained within one grade compared to professional evaluations, the system achieved accuracy of 98.90% for knitted fabrics and 98.19% for woven fabrics, which effectively verified the accuracy of this system in fabric pilling grading.
Conclusion The system is used for assessing the pilling grades of knitted and woven fabrics made from cotton and polyester materials. Through four evaluation indexes, the angle of incident light is determined to be 60°. Through the Spearman and Kendall correlation analysis, four characteristic parameters which have the greatest influence on the grade evaluation are determined. Among the machine learning classification algorithms, the random forest classification algorithm has the highest rating accuracy for knitted and woven fabrics, reaching 97.42% and 95.31%. By analyzing the range of characteristic parameters of different grades, the grading accuracy of knitted and woven fabrics can reach 98.90% and 98.19%. The final results verify that the system has high accuracy in the evaluation of fabric pilling grade. Future research plans involve further expanding the types of fabrics to include nonwoven fabrics. Additionally, other methods will be explored in the grading evaluation module to further enhance the accuracy of the evaluation results.

Key words: fabric pilling, pilling grading, neural network, correlation analysis, pilling characteristic parameter, machine learning classification algorithm