The application of automatic image recognition techniques in garment production (particularly for defect detection) is not new, although it has faced a range of limitations. In earlier stages, the process relied on manually extracting features from image pixel values, using statistical, spectral or structural methods. However, these approaches are sensitive to factors like lighting conditions, image background and the presence of patterns, among others. In the study Enhancing operational performance in textile manufacturing: impact of deep learning-based defect detection, researchers Artur Carvalho and Vera Migueis explore the integration of deep learning techniques and, above all, their impact from a productivity outlook.
Until now, beyond the various factors affecting detection, industries have also been heavily dependent on specialised operator expertise. In the case of so-called “warp defects”, operators must identify the issue on the production line, halt operations and resolve it. Until that happens, defective material continues to be produced and wasted. This leads to “limited reproducibility and adaptability across different domains”, which is particularly restrictive given the vast range of fabrics used in the textile industry and the many types of defects that may occur.
“This research surfaced within the scope of the PRODUTECH R3 programme,” explained Artur Carvalho, who is also a researcher at the Faculty of Engineering of the University of Porto. “The work aimed to reduce waste in the production process at IDEPA, a company based in São João da Madeira specialising in woven and printed labels, ribbons and other accessories.”
The use of deep learning techniques for defect detection, such as those proposed in the paper, has proven to be a “valuable alternative”, even prompting what can be described as a “revolution”. “These techniques, based on neural networks, enable the creation of end-to-end models that take pre-processed original images directly as input, thereby avoiding the specialised and time-consuming process of manual feature extraction,” said Artur Carvalho. He also mentioned that “deep learning architectures for image recognition are predominantly based on convolutional neural networks”, mechanisms capable of “acquiring the spatial structure of images”.
One of the standout aspects of the research lies in the application of established convolutional deep learning architectures both for image classification and for defect localisation – even in fabrics with multiple colours or unpredictable patterns. In the case of the company involved, which produces customised labels adapted to client needs, the need to “study these techniques in more sophisticated fabrics” also became clear.
Focus on operational impact
Beyond the technological dimension, the research introduces a new perspective by focusing on the operational impact of automated defect detection solutions, setting it apart from other case studies. According to Artur Carvalho, “more important than the predictive performance of the models is estimating their impact on metrics that matter to industrial managers, namely costs, actual production rates and defect rates”. He added that “this connection between predictive modelling and operational impact is not common practice in the literature in this field”.
Among the results achieved – most notably the “excellent predictive performance of the defect detection model” – it is also important to highlight the estimated reduction in monthly costs associated with identifying and locating defects. According to the data presented, between “92% and 98% of defective material could be avoided in the case of warp defects”.
News, current topics, curiosities and so much more about INESC TEC and its community!
