Abstract:

Objective The appearance quality of yarn is directly related to its mechanical properties and even economic value. However, manual inspection is still the dominant method in most factories, which are lagging and subjective. Based on machine vision and other emerging technology, in the ring spinning yarn production process of online detection of fine yarn evenness, hairiness and other indicators, so as to drive the classification of yarn drop and other emerging industry, has important theoretical significance and engineering value.
Method Accurate contour extraction during online visual inspection of ring spun yarn is difficult because of high speed yarn rotation and interweaving of the hairiness. To solve this problem, a method that fused deep-learning with morphological operations is proposed. Firstly, an online image acquisition system and focusing method are designed to provide high quality input for contour feature extraction; Secondly, a model based on holistically-nested edge detection(HED) neural network and morphological operations is constructed to achieve accurate online contour extraction under the interference of hairiness.
Results The camera was deployed to acquire 1 600 images of the yarn, whose resolution is 2 448 pixel×2 048 pixel, to calculate the optimal focal plane with the acquisition parameters. Compared with the images acquired under other focal plane parameters, the images acquired under the calculated focal plane parameters are of higher quality, which obviously improves the accuracy of contour extraction. 500 images of yarn were collected using the calibrated image acquisition system and processed with the proposed contour extraction method and other SOTA(state of the art) methods. The proposed method achieves OIS-F (optimal image scale), ODS-F (Optimal Dataset scale) of 0.91 and AP (average precision) of 0.89, which is more than 7% better than the current method. From the visual comparison results, it can be seen that the proposed method is based on the output of HED network, combined with morphological operations to remove the interference of hairiness and fiber texture, and reconstruct the yarn stem contour based on Cubic spline interpolation with good consistency. Finally, the extracted yarn contours were further processed using the proposed reconstruction method to calculate the CV value of yarn unevenness. In this paper, five groups of image data collected from different groups are processed using the proposed algorithm (experiments are performed using a Tesla V100 with 32 GB video memory GPU) to calculate the CV values for each group of 4 000 images. The average processing speed is about 24 frame/s, higher than the current experimental maximum image acquisition frequency of 20 frame/s. As shown in Fig. 6, the calculated results were compared with those of the laboratory high-precision electronic yarn evenness tester (CT3000), with an average error of less than 4% and a minimum accuracy of 92% and a maximum of 99% for a single group of tube yarn measurements.
Conclusion The image acquisition system calibration method improves the quality of the acquired data and facilitates the processing of subsequent algorithms. The designed deep learning and morphological operations fusion method for the extraction and reconstruction of yarn evenness effectively removes the interference of hairiness and improves the accuracy of the calculated CV values. In terms of processing speed, the proposed method can meet the current demand of online detection. And from the comparison results of the CV value of yarn unevenness, the detection accuracy also reaches the standard of practical application. The good application of the proposed method in the online detection of ring-spun-yarn evenness has been verified and the hardware system design as well as algorithm optimization can be further investigated for different application scenarios.

Key words: yarn evenness, online detection, machine vision, contour extraction, evenness CV values