Textile patterns are woven into the fabric of many parts of manufacturing, from clothes and soft furnishings to the interiors of luxury cars and public transport vehicles and beyond. A new approach to classification of these patterns based on artificial intelligence is discussed in the International Journal of Information and Communication Technology. ZhaoJue Dai of Wenzhou Polytechnic in Wenzhou, China, has developed an advanced method to automate textile pattern classification, which can cope with the incredible diversity of fabric designs.
Textile classification has traditionally been done by eye. But, in an era of information overload where there are myriad fabric designs entering the marketplace every day and patterns have become increasingly sophisticated as design, production technology, and dyes advance, classification needs more than an expert eye. Dai explains that computer vision, a branch of artificial intelligence that enables machines to see and interpret visual information, could solve the problem of textile overload.
Dai has now used convolutional neural networks (CNNs) to bring the process of textile classification into the digital age. She uses two techniques: mixture enhancement and attribute clustering within the analysis. Mixture enhancement can combine several textile images to create “enriched” digital swatches that can be used to train the CNN. This essentially teaches the computer to recognize novel patterns, improving its ability to handle the sheer diversity of textile designs in the real world. Attribute clustering then organizes the patterns by grouping together shared features. When presented with samples, the algorithm then has the ability to spot the nuances in a textile and classify it accordingly.
To fine-tune the process, Dai used entropy discretization. This technique converts continuous data into chunks that can be handled by the computer more efficiently as it compares different textiles. The system thus achieves a classification accuracy of well over 90 percent. This is better than previous textile classification models, which often unravel when presented with highly detailed or ornate designs.
Dai, ZJ. (2024) ‘Textile pattern style classification based on popular mixture enhancement and attribute clustering’, Int. J. Information and Communication Technology, Vol. 25, No. 8, pp.49–63.