基于形状语义的外观设计专利检索

Design Patent Retrieval Based on Shape Semantics

为跨越底层特征与高层语义之间的"鸿沟",充分挖掘外观设计专利图像形状组织的特点,实现了一个基于形状语义的外观设计专利检索系统.系统根据外观专利图像特点进行形状的预分类,利用专利形状类别之间的判别性信息提取形状语义,据此进行相似度匹配并返回相似专利图像.提出基于形状语义的相关性反馈方法,在初次检索结果的基础上,利用用户反馈图像的所属类别概率对相似度距离进行调整,进一步提高检索的精度.用实际使用的外观设计专利数据进行实验,结果表明,该方法取得了较好的检索效果,可以满足具有较高查全率的同时仍有较高查准率这一实际专利检索需求.

In order to improve the performance of design patents retrieval, design patent image retrieval system based on shape semantics is proposed to exploit characteristics of shape organization of design patents to bring the challenging ＂semantic gap＂. Design patent images are classified according to the shapes, and discriminative information between different patent shape categories is utilized to extract shape semantic features. Then the similarities of semantic features between query image and images in database are calculated and sorted to obtain the most similar design patents. A shape semantics based relevance feedback is proposed to improve retrieval results by adjusting feature distances based on the initial retrieval results and category information of feedback images. Experiments on real design patent images demonstrate the effectiveness of proposed method. The retrieval system meets the practical design patent retrieval demand which should reach high precision on high recall.