A knowledge augmented deep learning method for vision-based yarn contour detection. (April 2022)

• Record Type: Journal Article
• Authors: Xu, Chuqiao; Wang, Junliang; Tao, Jing; Zhang, Jie; Zheng, Pai
• Is Part Of: Journal of manufacturing systems. Volume 63(2022)
• Page Start: 317
• Page End: 328
• Publication Date: 2022-04
• Subjects: Knowledge augmented -- Deep learning -- Contour detection -- Machine vision -- Quality inspection -- Yarn manufacturing; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2022.04.006
• Languages: English
• ISSNs: 0278-6125

Abstract

Abstract: Contour detection extracts accurate object boundaries from natural images, which plays a critical role in the vision-based inspection. However, contour detection is still challenging in yarn quality inspection tasks, since the current methods pursue global high-precision contours while ignoring detecting the boundaries of interest from intertwined objects in complex local regions. This paper proposes a knowledge augmented contour detection method with deep learning to extract pure backbone boundaries from intertwined objects for more accurate yarn quality inspection. Firstly, according to the directional gradient distribution of yarn images in the visual system, a radial pixel difference convolution is improved to extract interested edge features. Secondly, considering the graphical characteristics of the yarn to be inspected, a mask layer of texture erosion is designed to further filter irrelevant details from extracted edge features. The comparative experiments during actual spinning processes demonstrate that the proposed method achieves a better ODS (optimal dataset scale, a standard contour detection evaluation measure) of 0.815, and reduces the quality inspection error from about 1.5% to tolerated 0.534%. Highlights: Yarn evenness inspection is closely related to edges of interest rather than global contours. The hairiness and texture of the yarn are sensitive to the quality inspection accuracy. Injecting the prior knowledge augments the detection ability of contour detection neural networks.