目的 针对实际生产中获取印刷标志图像标签成本较高的问题，研究基于约束谱聚类的印刷套准状态识别方法。方法 基于少量有标签的样本，建立样本之间的must-link约束和cannot-link约束，并进行约束扩展。计算印刷标志图像样本点欧式空间相似度矩阵，并根据扩展后约束关系修正，构建样本点的特征向量空间。采用K-means方法对样本点特征向量空间进行2类聚类，即印刷套准图像和印刷套不准图像。结果 文中方法在实验数据集上的最高印刷套准识别准确率为98.11%。文中方法(约束对数为30)的识别准确率优于无监督的谱聚类方法、朴素贝叶斯方法和决策树方法，文中方法与SVM方法的识别准确率接近。文中方法获取印刷标志图像标签的成本低于SVM方法，且模型建立和识别的时间也少于SVM方法。结论 文中方法以较少的获取印刷标志图像标签成本达到了较高的印刷套准识别准确率。

The work aims to study the printing registration recognition method based on the constrained spectral clustering algorithm, in order to overcome the problem of high cost of obtaining the label of the printing mark images in actual production.The constraints of the must-link and the connot-link were set and extended based on a few of labeled samples. The similarity matrix of the samples was computed in the Euclidean space and then revised by the extended constraint relationship, to construct the eigenvector space of the samples. The K-means method divided the samples in the eigenvector space into two sections, including the samples with the registration label and the ones with the misregistraion label. The proposed method was performed the experimental data, with the best printing registration recognition accuracy of 98.11%. The proposed method with 30 pairs of constraints outperformed the unsupervised spectral clustering algorithm, the Naivebayes method and the decision tree method in terms of the recognition accuracy. The recognition accuracy of the proposed method was nearly equal to the one of the support vector machine (SVM) method. Besides, compared with the SVM method, the proposed method was lower in the cost of obtaining the labels of the printing mark images and in the time-consumption of model construction and recognition.The proposed method achieves the higher accuracy of printing registration with the low cost of acquiring the labels of printing mark images.