融合金字塔和注意力机制的文物子图检索模型

A cultural relic sub-image retrieval algorithm based on folded multi-hollow pyramid pooling and attention mechanism

随着中国文化研究工作的深入以及数字化文物采集技术的发展,文化资源数据和文化数字内容的数量也随之增长,如何对文化数据进行有效存储、管理以及检索成为一项重要的工作。针对文物图像数据检索任务中因尺度变化和特征选择造成检索精度不高的问题,提出了一种融合折叠多空洞金字塔池化和注意力机制的文物子图检索模型。模型为提高不同尺度的文物子图检索精度,通过在图像特征提取模块使用优化后的折叠多空洞金字塔池化提取图像的多尺度信息;为避免密集局部特征和无关特征影响检索准确率,使用注意力机制对局部特征进行关键特征选择。最后在所构建的文物数据集上进行了消融实验和性能对比实验,实验结果取得了良好的效果,mAP达到85.3%。

With the deepening of research on Chinese culture and the development of digital cultural relics collection technology,the amount of cultural resource data and cultural digital content has also increased,so how to store,manage and retrieve cultural data has become an important task.In order to solve the problem of low retrieval accuracy caused by scale change and feature selection in cultural relic image retrieval tasks,in this paper a cultural relic sub-image retrieval algorithm based on folded multi-hollow pyramid pooling and attention mechanism(FMHPPA)was proposed.In order to solve the problem of scale change in sub-image retrieval,FMHPPA model extracted multi-scale information from image feature extraction module by optimizing folded multi-hollow pyramid pooling.In order to avoid the impact of dense local features and irrelevant features on retrieval performance and accuracy,FMHPPAM model used attention mechanism to select key features for local features.The model ablation experiment and performance comparison experiment were carried out on the constructed sub-image dataset,and the experimental results achieved better results as the mAP reached 85.3%.