基于对比学习的无监督商标检索

Unsupervised trademark retrieval based on contrastive learning

商标图形要素是商标审查时,审查人员判断两个商标是否相似的重要判断依据.然而由于商标图形要素具有多种可能的解释,不同审查人员对同一商标图形要素可能有不同的理解,造成商标图像具有丰富的语义信息.对比学习通过最大化正样本语义信息来训练模型,能根据商标图像含有的独特语义信息对其进行区分,有助于完成商标检索等下游任务.为了充分理解商标图像的语义信息,首次将无监督对比学习引入到商标检索领域.首先使用一系列能增强语义信息的预处理方法进行处理,从而保留商标图像所含有的语义信息;然后根据商标图像生成的正负样本视图将数据进行分组,着重学习到正样本之间的相同语义信息以及正负样本之间的差异信息;所提取到的语义信息有助于模型判别两张商标图像是否语义相似,最终提高商标检索的平均精度.在公开的百万级商标图像数据集METU上进行的测试结果表明,与现有最新的商标检索方法相比,本文所使用的对比学习算法将mAP@100指标从55.0%提高到72.6%.

The graphic element of a trademark is an important basis for the examiners to judge whether two trademarks are similar during trademark examination.However,due to the many possible interpretations of trademark graphic elements,different examiners may have different understandings of the same trademark graphic elements,resulting in rich semantic information in trademark images.Contrastive learning trains the model by maximizing the semantic information of positive samples,which can distinguish trademark images according to their unique semantic information,which is helpful for downstream tasks such as trademark retrieval.In order to fully understand the semantic information of trademark images,for the first time,introduced contrastive learning into the field of trademark retrieval.Firstly,a series of preprocessing methods that can enhance semantic information are used to preserve the semantic information contained in the trademark image;Then group the data according to the positive and negative sample views generated by the trademark image,and focus on learning the same semantic information between positive samples and the difference information between positive and negative samples;The extracted semantic information helps the model to distinguish whether the two trademark images are semantically similar,and ultimately improves the average accuracy of trademark retrieval.The test results on METU,a public million-level trademark dataset,show that compared with the existing state-of-the-art trademark retrieval methods,the proposed algorithm improves the mAP@100 from 55.0%to 72.6%.