关键词: 刺绣针法, 针法模型, HC-Snake模型, 灰度共生矩阵, 颜色矩, BP神经网络

Abstract: The traditional embroidery is the intangible cultural heritage of the Chinese nation. and is vigorously promoted by the state. At present, most studies focus on the cultural heritage of embroidery and the application of embroidery in fashion products, while research for embroidery stitches is rarely found. In particular, the research on the recognition of embroidery stitches from the perspective of images is even less. Recognition for embroidery stitches is mostly done by experience, while few methods are available for intelligent recognition, which is not conducive to the digital protection of embroidery works.
In order to realize the intelligent recognition of embroidery stitches and promote the digital protection of embroidery, we firstly classify embroidery stitches as three classifications: point stitches, linear stitches and block stitches on the basis of analyzing the characteristics of embroidery stitches. Block stitches are the stitches whose proportion is the highest among the three classifications in the same embroidery works, while the flush stitches, the grabbing stitches and the overlapping stitches are three kinds of basic stitches which block stitches contain. With the three basic stitches as the processing objects, we analyze and compare the characteristics of these stitches, and construct the stitch model.
Secondly, the texture details of the embroidery stitch image are enhanced, and on this basis, the traditional Snake algorithm is improved. Based on Harris corner detection, corner points in the image are calculated to form a corner set. The Canny operator is used to detect the edge in the image, and the point set of the initial target contour is selected. Then, with the initial target contour as the center, the 5*5 template is used to cover the corner points, and the contour control points of Snake algorithm are determined, forming an intelligent HC-Snake model to realize the intelligent recognition of the target contour. The texture and color features of the target image including a total of 12 feature parameters such as the mean and standard deviation of contrast (CON), homogeneity (IDM), energy (E), entropy (ENT) and correlation (COR), the second-order color moments (SM) and third-order color moments (TM) are extracted by using the gray level co-occurrence matrix and color moments. Through the analysis of variance of the data, the significant degree of the influence of the stitch classification on each index is verified. It is found that the nine characteristic parameters including CONsd, IDMsd, Ea, Esd, ENTa, ENTsd, CORsd, SM and TM are significantly affected by the stitch classification, and they are identified as the characteristic indexes for classification, and a characteristic data set is established.
Finally, the BP neural network classification model is established, and the parameters of the neural network model are determined by comparative analysis, and the needle samples are classified and identified. With the help of confusion matrix, the accuracy and recall of the classification model are calculated to confirm the reliability of the model. In the meanwhile, the recognition model is verified by an example. The experimental results show that the model can realize the classification of embroidery stitch images, and the classification accuracy reaches 93.3%. 
The HC-Snake algorithm proposed in this paper can effectively improve the shortcomings of the traditional Snake algorithm in manually selecting control points, and make it more intelligent. The establishment of embroidery stitch recognition models is used for stitch recognition of ordinary network embroidery images, which improves the accuracy and applicability of the recognition model. The research results can be used to promote the digital protection of traditional embroidery.

Key words:  , embroidery stitches, stitch model, HC-Snake model, gray level mo-occurrence matrix, color moments, BP neural network