一种面向商品检索的多尺度度量学习方法
A Multi-scale Metric Learning Approach for Product Retrieval
行阳阳 1张索非 1宋越 2吴晓富 1周全1
作者信息
- 1. 南京邮电大学,江苏 南京 210003
- 2. 95958 部队,上海 200120
- 折叠
摘要
商品图像检索是一个典型的大规模度量学习任务,其特点在于商品零售平台需要定期上架新类型的商品,且同一类型的商品外观会不时发生变化.已有的工作表明:传统基于单一的度量学习虽然可以将商品检索模型的识别范围扩展到未知商品类别上,但是其性能仍然受限.为此,提出了一种基于多尺度监督信息的深度度量学习商品检索方法.该方法利用商品多个尺度的标签信息训练并使用协同注意力机制对不同尺度的深度特征进行有效融合,提高了深度学习模型挖掘重要信息的能力,从而有效提高了其在细粒度级别下的检索性能.在大规模商品检索数据集上的实验结果表明,该方法在mAP和Rank-1 上分别为43.0%和65.9%.相比于传统度量学习方法分别提升了6.4%和7.8%.
Abstract
Product image retrieval is a typical large-scale metric learning task.The specificity of this task is that the commodity retail platform needs to import new types of items regularly,and the appearances of existing products also change from time to time.The previous work show that although the traditional metric learning can extend the recognition range of product retrieval to unseen product types,the performance of traditional metric learning in product retrieval is still limited,because it can only use a single scale of regulatory information to deal with such large-scale retrieval problems.Therefore,we propose a product retrieval method based on multi-scale deep metric learning.The proposed method uses the label information of multiple scales to train the model and adopts the co-attention module to integrate the deep features of different scales effectively,which can improve the ability of the model to obtain important information and effectively improve the retrieval performance of the deep learning model at the fine-grained level.The proposed method achieves 43.0%and 65.9%on mAP and Rank-1 in experiments on large-scale product retrieval dataset,improved by 6.4%and 7.8%,respectively compared with the traditional metric learning.
关键词
度量学习/商品识别/多尺度/图像检索/特征融合Key words
metric learning/product identification/multiple scale/image retrieval/feature fusion引用本文复制引用
基金项目
国家自然科学基金面上项目(61876093)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据