中国传统服饰作为中华文化重要的符号，图像内容丰富，在图像分类领域有着巨大的潜在价值。针对图像分类中传统判别字典学习方法存在的原子局部特征利用不足且单层结构在多标签分类上的局限性，提出了一种结合全局约束和支持向量引导的多层字典学习算法，实现单层到多层的过渡来获取高层复杂语义信息。该算法通过引入全局约束增强非相邻原子间相似性的考量，更精准地捕捉图像特征的非线性关系；同时采用支持向量判别项优化编码向量对的权重分配，提高模型的泛化能力。实验结果显示，此算法在明清服饰纹样数据集上相较于局部约束和单层字典学习方法，分类精度分别提高了 3.33%和 1.57%，在扩展耶鲁B人脸数据集上也展示了较高的分类准确度

As an important symbol of Chinese culture, traditional Chinese clothing has rich image content and great potential value in the field of image classification. Aiming at the insufficient utilization of atomic local features in traditional discriminant dictionary learning methods in image classification and the limitations of single-layer structure in multi-label classification, a multi-layer dictionary learning algorithm combining global constraints and support vector guidance was proposed to achieve the transition from single layer to multi-layer to obtain high-level complex semantic information. The algorithm introduced global constraints to enhance the consideration of similarity between non-adjacent atoms and more accurately captured the nonlinear relationship of image features; at the same time, the support vector discriminant term was used to optimize the weight distribution of encoding vector pairs to improve the generalization ability of the model. Experimental results show that compared with local constraints and single-layer dictionary learning methods, the classification accuracy of this algorithm on the Ming and Qing clothing pattern dataset is improved by 3.33% and 1.57% respectively, and it also shows high classification accuracy on the extended Yale B face dataset