针对有效、准确地检索传统纹样子图数据的原图的问题，提出了一种基于Transformer的传统纹样子图检索算法。首先构建传统纹样子图数据集，其次利用卷积神经网络提取多层次的特征图进行融合。针对数据库中的图像利用Transformer生成预测框，将预测框映射回融合特征图提取局部与全局特征图后利用特征聚合算法聚合为全局与局部特征向量；针对查询子图，仅利用特征聚合算法对融合特征图进行聚合为全局向量。之后将查询子图的特征向量与数据库图像的特征向量分别进行相似度计算，排序后得到检索结果。

Aiming at the problem of retrieving the original image of traditional pattern sub-image effectively and accurately, a traditional pattern sub-image retrieval algorithm based on Transformer is proposed. Firstly, the traditional sub-image dataset is constructed, and then the convolutional neural network is used to extract the multi-level feature for fusion. For the image in the database, Transformer is used to generate the prediction box, and the prediction box is mapped back to fusion feature images to extract local and global feature images, and then global and local feature vectors are aggregated by feature aggregation algorithm. For the query sub-image, the feature aggregation algorithm is used to aggregate the fusion feature graph into global vector. After that, the feature vectors of the query subgraph and the database image are respectively calculated for similarity, and the retrieval results are obtained after sorting.