Image Retrieval Based On Spatial Relationships Similarity

With the rapid development of Internet technology, the information that users processdaily is going to be massive and to be diversified. On the web, there are more and moreimages viewed as context. Thus, image retrieval has become one of the most popularretrievals. The traditional search engines on the Internet, such as Google, Yahoo!, and Bing,provide image search functions. However these functions are based on the file name, or on thekeywords embedded in the web page around the image. The gap between understanding ofimages and texts expression, always leads to the decrease of precision rate, users can’t obtainthe wanted results even through the search engines recommend lots of results. At thiscircumstance, researchers provide content-based image retrieval (CBIR). CBIR means that thequery itself is an image or a description of the image content, we can contribute index byextracting the features, and then compute the distance between the database images’ featuresand the query image’s features so as to measure the similarity.CBIR based on spatial relationships is one of the branches in the CBIR based onsemantic features. After obtaining the symbioses of objects in images, first we should get thecommon objects appeared in both images, and then estimate whether or not these objects canbe the same spatial relationships. The objects are considered as“similarity objects” if and onlyif the relationships between the two objects in query image and database image are identical.Sometimes, users can’t provide a perfect image as a query one. The query is justsimilar to the wanted image or just similar to their understanding of their intention, at thiscircumstance, after offering users the recommend results, the retrieval system must feedbackseveral times. Among the recommend results users mark some relevant images as a positiveset, the set expresses some available information can be added into the query features. In thisway, we refine the query in order to be closer to the intention. In this paper, we have two partresearch issue:First of all, this paper studies some image retrieval based on spatial relationships andmakes some improvement on CPM, proposed a new similarity retrieval by rectangle algebra,this method takes the advantage of MBR present an object and use rectangle algebra toestimate the spatial relationships between objects. Compared to type-i in CPM, MBR can bemore precious in terms of expression of the objects’ range and location, what’s more,rectangle algebra is more accurate, so the measure results are of great nicety. During theprocess of searching, the proposed algorithm SRRA prunes lots of redundant information,making contribution to shorten comparison time. Second, this paper uses Query Reformulation for relevant feedback. After each round,users should select some relevant images as positive examples, and the default ones areviewed as negative ones. We bring available extra information into previous query and takeneedless information away, in this way we reformulate query and search again. In this paper,we proposed a new relevant feedback algorithm based on SRRA. This method uses thesymbolic expressions of positive examples to compute maximum sequential pattern,maximum sequential pattern includes the common information which can be brought into theoriginal query and make some adjustment towards the original query, uses SCS to put themaximum sequential pattern and query together as a new string presenting a new query image,and the third step is to confirm spatial relationships between any objects pair. After the threesteps, we reformulate the query features, and can search images again. When the steps offeedback archives system requirement or users stop the process, the retrieval ends.The experiments include two parts, and one dataset is artificial, the other comes fromhttp://wang.ist.psu.edu/docs/related/. The results show that:1. Compared to CPM, the approved SRRA algorithm cannot only provide more preciserecommend results and also shorten retrieval time nearly50%;2. According to relevant feedback algorithm, combining the original query image and theavailable information, we can describe the features more precisely, so these recommendimages can be more fit for users’intention.