面向检索应用的商标显著性检测方法

Trademark Saliency Detection Method for Image Retrieval

商标显著性检测是商标检索的重要前提之一,待申请商标需要商标显著性判断,商标相似或侵权判定也需要对商标显著性特征加以判别。考虑到商标多由图案组成,该文提出了一套面向商标图像的显著性检测方案。首先,基于中国商标数据库内的商标图像抽取、加工并制成商标数据集,搭建商标数据库并陆续开源一批商标显著性检测数据集。基于已有显著性检测框架,开发并评估了多种主流显著性检测算法。结果表明一种适配商标图像的U 2-Net深度模型对商标显著性检测效果较好,综合准确率在92%左右,后续还需要深入优化和评测。最后,提出一个面向相似商标检索的显著性检测服务和特征生成解决方案,并开发了相关搜索系统,为后续工业级应用奠定基础。

Trademark saliency detection is one of the important prerequisites for trademark image retrieval.The trademark should be applied and registered after the saliency judgment.For the determination of trademark similarity or infringement,it is also necessary to distinguish saliency features of trademarks.We propose a trademark image-oriented saliency detection scheme.Firstly,in view of the lack of related datasets,a trademark database was built based on the trademark images which are extracted,processed from the Chinese trademark database.After a further evaluation,an open-source trademark saliency detection dataset is proposed.Based on the current saliency detection framework,a variety of mainstream saliency detection algorithms have been developed and evaluated.The results show that a U 2-Net network adapted to the trademark images has the most outstanding performance on the trademark saliency detection,and the overall accuracy is about 92%.The further model optimization training and evaluation should be later established.Finally,a module integrated with the saliency detection service and the feature generation service for the trademark retrieval is proposed.An image-to-image trademark retrieval system has been developed,which lays a foundation for the subsequent industrial applications.