基于多层次特征的跨场景服装检索

Cross-scenario clothing retrieval based on multi-level features

针对跨场景服装检索如何提取更有表述力的服装特征问题,提出了一种新的基于高层公共特征约束的相似性度量算法。首先,通过类别空间学习提取不同场景域的类别信息;然后,在场景域空间网络用类别信息约束传统对比损失函数,增大对类间负样本对的惩罚以减轻过拟合;最后,融合类别公共特征和域特定特征并通过类别判断进行辅助检索。分析和实验结果表明,新算法对跨场景服装检索的准确度要优于当前前沿的方法。

The current cross-scenario clothing retrieval framework is based on expressive feature extraction. A new similarity measure based on category constraint is proposed. First, the category information of different scene domains is extracted through category space learning;The category information is then used to constrain the traditional contrast loss function in the scene domain space network. By this way, the penalty of the negative sample pair is increased to alleviate the over-fitting. Finally, we combine the common category features with domain-specific features to carry on retrieval with category constraint. The analysis and experimental results show that the new algorithm outperforms stateof- the-art methods in terms of cross-scenario clothing retrieval.