A two-scale attention model for intelligent evaluation of yarn surface qualities with computer vision

Abstract

In this paper, an intelligent computer method for yarn surface grading is developed to analyze yarn board image and objectively evaluate yarn quality according to ASTM D2255.Both statistical features measuring the overall performance and relative features measuring salient regions are elaborately designed and selected. Statistical features are extracted to characterize the yarn body and hairiness. In relative feature extraction, a two-scale attention model is proposed and developed, which can fully imitate human attention at different observation distances for the whole and detailed yarn information. Global and individual Probabilistic Neural Networks (PNNs) are then designed for yarn quality evaluation based on eight-grade and five-grade classifications. A database, covering eight yarn densities (Ne7~ Ne80) and different surface qualities, was constructed with 296 yarn board images for the evaluation. The accuracy for eight- and five-grade global PNNs are 92.23 and 93.58%, respectively, demonstrating a good classification performance of the proposed method.

Keywords:

• Yarn evaluation
• feature extraction
• computational attention model
• DPI determination
• probabilistic neural network
• classification

Acknowledgments

Dr. Li SY would also thank the Hong Kong Polytechnic University for providing her with postgraduate scholarship.