基于特征的文档子图像检索及其相关反馈

Feature-Based Document Sub-Image Retrieval using Relevance Feedback

探讨纯文本图像的子图像检索问题。提取其层次结构特征进行匹配,同时为了提高检索精度,又提出了一种适合文档子图像检索的相关反馈算法。实验采用6千幅英文手写体纯文本图像作为样本集,每次迭代返回给用户12幅图像,结果表明每次迭代用时约4秒,6次迭代后召回率基本稳定在83%。

For text image, its hierarchical characteristics is extracted to solve the problem of feature - based sub-image retrieval. To improve the retrieval precision, a suitable relevance feedback algorithm for document sub - image retrieval is presented ： for each iteration, positive and negative samples are specified by the user and for each sample image in the cumulative positive and negative samples the system updates the penalty coefficient of its sub- image and determines the query points moving to the positive and negative sample set, then the system updates the query image combining Rocchio algorithm. Performance evaluation on a dataset of over six thousands English handwritten text images shows that the system can achieve a stable average recall value of 83% with 6 iterations while for each such iteration the user can get the top 12 retrieved images in 4 seconds.