| As the growth of image data in the internet, It is getting harder for user to get useful information from the massive data. Image retrieval plays a more important part which helps to retrieval information effectively. One of the major obstacles of retrieval system is to understand user intents through the query they gave and properly provide results that meet their need. One reason that causes this is the ambiguity of users’ query; the other is the near duplicate results. Ambiguity will cause retrieval system to give wrong results while near duplicate will decrease the usefulness of retrieval results provided. The semantic gap of image and the increasing near duplicate images actually magnified this problem.Works in image retrieval had noticed the bad effect caused by too much near duplicate images. They try clustering and eliminating near duplicate to increase the diversity of retrieval results. Text retrieval, however, defines new ranking functions to include more subtopics related to users’ query in retrieval results, so as to solve ambiguity of user’s query. Noted that users’ intent vary for different user at different time even with the same query. The previous work tries to please every user with one criterion, ignoring different user’s preference. Thus people had involved user’s relevance feedback in image retrieval to grasp users’ intent.With near duplicate problem, we designed a zoom algorithm for image retrieval results. Users’ may directly control the diversity of retrieval results through interactive component provided:zoom factor. To solve the ambiguity problem, a semantic extraction algorithm is designed. It do a clustering for the retrieval results first considering images’ tags as their semantic information and then extract keywords for every cluster. Finally we design and implement a prototype system combining these two algorithms to show how to embed them into image retrieval systems, so as to provide a mechanism to analyze user intents for image retrieval system. |
PDF Full Text Request