基于支持向量机的筒子纱纱管品种检测

为了对运输导轨上筒子纱的纱管品种进行检测，提出一种基于纱管图像分类的品种检测方法。首先通过搭建的采集装置采集筒子纱顶部包含纱管信息的图像，运用阈值分割和椭圆拟合得到纱管区域，利用极坐标变换将纱管圆环展开成矩形图像，然后使用HSV颜色直方图和局部二值模式分别提取纱管展开图像的颜色特征和纹理特征，最后通过支持向量机构建筒子纱纱管品种分类模型实现纱管品种检测。采用建立的纱管品种检测分类数据集进行实验，结果表明，本文方法相比于其它特征组合和分类器，具有更高的分类准确率，对相同图案的星型纱管、黑色系花纹纱管和混合纱管的分类准确率达100%,可为纺纱企业筒子纱纱管品种检测与运输包装提供参考。

Detection of cheese yarn bobbin varieties based on support vector machine

Objective Aiming at variety detection of cheese yarn on the transport guide rail in practical situations, this paper proposed an automatic solution based on image processing. Based on the characteristics of bobbin, a yarn variety detection method based on bobbin image classification was proposed to replace subjective judgment. The method proposed in this paper is aimed to reduce error rate of manual detection and labor costs, and to improve the production efficiency of spinning.Method In order to facilitate feature extraction, the original image of cheese yarn was processed by the image segmentation method to obtain the bobbin area. Then, the segmented image of the annular bobbin was expanded into a square image by polar coordinate transformation. Color and texture features were extracted from the expanded image and optimized based on feature classification experiments and the elapsed time to jointly characterize the bobbin image.Results The Otsu threshold method was adopted to find the gray threshold of the bobbin foreground and background, and a binary image was obtained based on the determined threshold. The image contour was used to filter non-target regions in the binary image by setting the area and perimeter threshold of the bobbin region. The binary image containing only the region of the bobbin was adopted as a mask to segment the yarn tube from the original cheese yarn image. For the segmented annular yarn tube, the polar coordinate transformation was applied to transform it into a rectangular image with the circumference of the outer circle and the width of the outer circle. Bobbin image expanded after threshold segmentation provided data support for subsequent research. The non-uniform quantized color histogram features with H:S:V=8:3:3 were optimized by the classification accuracy and the elapsed time of feature extraction (Tab.1). The features obtained by the preferred color quantization method demonstrated satisfactory classification effect on different types of bobbin images with the acceptable calculation time. The performance of different local binary pattern operators was optimized by the bobbin classification experiments. The operator of rotation invariant equivalent LBP_16,2 with the sampling point of 16 and the sampling radius of 2 was optimized to extract the texture features of the bobbin image (Tab.2). The experiments using the fusion feature were performed on bobbin with the same color, the same pattern, different colors and patterns and the results proved that the fusion feature is able to adapt to the bobbin classification detection task in the three cases (Tab.3). The classification results of different feature combinations under the same classification model and the performance of fusion features on different classifiers proved that the method of combining color and texture and support vector machine classification model is effective for bobbin detection (Tab.4, Tab.5).Conclusion The combination of Otsu threshold method and image contour is successfully used to segment the bobbin from the original image, and it is shown that the bobbin image expanded by polar coordinate transformation can facilitate feature extraction. The optimal technique to characterize the color information of the bobbin image is proven to be the non-uniform quantization color histogram feature of H:S:V= 8:3:3. The preferred way to characterize the texture features of the bobbin image is the histogram feature of the rotation invariant equivalent LBP_16,2. The fusion of the two features can deal with the classification detection tasks of the bobbin in the same color, the same pattern, and different colors and patterns. Different classifiers were compared by the experiments and support vector machine was selected as the optimal classification model because of the best performance. The classification accuracy of this method is 100% in the same pattern of star pattern bobbin, black pattern bobbin and mixed pattern bobbin, which is the highest among different methods. The proposed method has practical value for variety detection of bobbin on the transport guide rail.