基于图案的多层次跨场景服装检索

Pattern-based Multi-level Cross-scene Clothing Retrieval

在跨场景服装图像检索问题上,目前的“以图搜图”方法主要依赖于服装款式类别等服装区域全局特征提取的结果,导致检索准确度不高。针对这种情况,提出一种基于图案的多层次跨场景服装检索方法,在考虑服装区域特征匹配的基础上,同时将服装内部图案区域特征的匹配度作为检索结果的衡量标准,使用多层次的服装特征描述进行检索匹配。此外在进行度量学习时使用triplet loss函数实现,将不同背景环境下的图像作为正样本输入训练网络可以提高网络模型对于跨场景服装检索的抗干扰能力。实验结果表明,图案区域匹配度的加入和triplet loss函数的使用可以有效增强网络的特征提取能力,能够提高跨场景服装检索的精度,适用于日常服装检索。

In the cross-scene clothing image retrieval,the current"searching pattern by pattern"mainly relies on clothing regions global feature extraction of garment styles,resulting in low retrieval accuracy.In view of this situation,this paper proposes a pattern-based multi-level cross-scene garment retrieval method.Based on the feature matching in the garment area,the matching degree of the regional patterns in the interior of the garment is taken as the measure of the retrieval result,using the multi-level garment feature description to match the search.In addition,the triplet loss function is used to measure learning,and inputting the images under different background environments into the training network as a positive sample can improve the anti-interference ability of the network model for cross-scene garment retrieving.The experimental results show that the addition of pattern region matching and the use of triplet loss function can effectively enhance the feature extraction ability of the network and improve the accuracy of cross-scene clothing retrieval,which is suitable for everyday clothing retrieval.