[4]. In general, denition of term weighting scheme is shown below. where is the weight of term in document is term count, is document count. Literature wij j i, n m proposes a plenty number of weighting schemes. We used pivoted unique term weighting scheme best performance among the other methods, we chosen to use C4.5 in training phase [3]. After all, We evaluated some machine learning algorithms to use in training phase. Table 2 shows results precision at top ranks. Re-ranking phase includes training. So, it rst requires to extract features from both documents and query [2]. Table 1 has a list of used features in two groups.
After generating initial results, re-ranking phase re-orders initially generated results to obtain high new ranks, of all documents in the initial results is recalculated using following formula: Rnew, of 10-fold cross-validation of each method. Because C4.5 Decision tree generation algorithm shown Feature Denition P (G) 1 PiN=i1 P (G) Ni P (C) N1i PiN=i1 P (C) a modication on matrix (Eq. 1) by adding visual terms representing visual contents. Formally, D document-term matrix becomes as follows: Our second approach focuses on closing semantic gap problem of content-based image retrieval.
Our approach aims to integrate both textual and visual contents in same space. System proposes combines visual features to traditional text-based VSM. Initially, we chosen to use two simple and textual features are normalized independently. In sum, Integrated Retrieval Model (IRM) where is the weight of image term in document is the number of visual terms. Visual iij j i, k of one term is given. Algorithm counts pixels that has same pixel value in each color channel. This visual terms which aim to model color information of the whole image. In Algorithm 1 extraction
⎡ w1,1 w1,2 ⋅ ⋅ ⋅ w1,n i1,n+1 i1,n+2 ⋅ ⋅ ⋅ i1,n+k ⎤ D′ = ⎣⎢⎢⎢ w...2,1 w...2,2 ⋅. ⋅. .⋅ w...2,n i1,...n+1 i2,...n+2 ⋅. ⋅. .⋅ i2,...n+k ⎦⎥⎥⎥
wm,1 wm,2 ⋅ ⋅ ⋅ wm,n im,n+1 im,n+2 ⋅ ⋅ ⋅ im,n+k 3 and Table 4 respectively. According to imageCLEFmed 2008 results our re-ranking approach we used ground truth data of imageCLEFmed 2008 data, system learns relevance information of 6 %. Reason of this loss lies in the dierence between training and evaluation datasets. Since For example, dash(-) character is removed because of the importance of terms like x-ray, T3-MR about IRM. Before going further, we performed a preprocessing in all 74902 documents including method is used. ImageCLEFmed 2008 dataset also contains 30 queries and their relevance data. imageCLEFmed 2008 dataset only. boosts performance of both methods. However, results of imageCLEFmed 2009 did not show same In this section, we describe experimentations performed in ImageCLEFmed 2009. Our experimedical articles. Number of index terms is 30343 and same term generation technique with baseline like dash(-) and apostrophe(’) is removed, others like comma(,), slash(/) is replaced with a space.
Experimentations of Re-Ranking method includes training phase where we use imageCLEFmed These 30 queries categorized under 3 groups; visual queries which results will be better by using These index terms are normalized as described in Eq. 3. We evaluated performance of this baseline mentations can be classied into two groups. The rst one is about re-ranking and second one is and textual queries which targets to textual retrieval systems. Experimentation results of our 2008 data. Training data contains titles and gure captions of approximately 67000 images in uniformity. Numbers in the documents are not removed. However, some punctuation characters etc. We chosen words surrounded by spaces as index terms. Finally we had 33613 index terms. re-ranking approach on both base-line and IRM on imgeCLEFmed 2008 data is given in Table visual content, mixed queries which are prepared to test performance of mixed retrieval systems method and Figure 2 shows results of this run. combination of title and captions. First, all documents were converted to lower-case to achieve impact. According to Table 5, proposed re-ranking technique reduces system performance about 0.352 0.288 0.425 0.263 bpref 0.278 0.368 0.200 0.385 methods shows the best performance in all measures. model on ImageCLEFmed 2008 dataset. Whereas our model shows similar performance with Integrated retrieval model experimentation needs image features to be extracted rst. As of our integrated model can be improved by using re-ranking algorithm and combination of two baseline method on visual queries, it shows better performance on other two query types. Results mentioned before, we used simple grayscaleness feature. Table 4 presents results of our integrated is given. According to the gure our IRM outperforms all of our runs in means of recall at all We participated ImageCLEFmed 2009 with 5 ocial and we evaluated our re-ranking approach cording to the gure, our re-ranking method increases recall levels when it is applied to both of the to the ImageCLEFmed 2009 results. In Figure 2 Precision and recall graph of all our methods CLEFmed 2009 results. In Table 6 ocial results of automatic mixed retrieval area is given. Our all cases. feature is used as a visual term, performance of the approach will be expected to improve with than baseline method, Re-Ranking performance of mixed retrieval shows the best recall levels in approaches, base-line and integrated retrieval. Since results of integrated retrieval model is better on IRM run unocially . In Table 5, performance of all our methods is given on ImageCLEFmed Our Integrated VSM approach has best scores in automatic mixed retrieval area of the ImageIntegreated VSM technique has the highest MAP score of Mixed Automatic runs. precision levels. new features. As mentioned before our re-ranking approach reduces system performance according evaluated methods based on recall-precision scale in imageCLEF2008med dataset is given. AcIn meaning of precision our methods outperforms baseline. In Figure 1 performance of all others. Only Re-Ranked results of IRM run shows same performance at P@5 scale. Since simple 2009 dataset. Our Integrated VSM approach shows best performance in all measures among 0.560 0.440 0.420 0.340 0.320 0.313 P@5 0.473 0.280 0.089 P@500 0.093 0.082 0.107 0.088 0.089 0.096 0.092 0.211 0.259 0.268 0.312 0.384 MAP 0.160 0.130 0.190 2008 data.
Table 3: Performance of Our Re-Ranking Approach on Base-Line approach using ImageCLEFmed 1553 848 1757 1722 1659 1176 1719 1754 1724 RelRet 1763 York University_79_8_1244834740798 York University_79_8_1244834388965 ITI_26_8_1244811028909 DEU IRM York University_79_8_1244834823312 York University_79_8_1244834655420 medGIFT_77_8_1244842980151 University of North Texas_55_8_1244879759190 medGIFT_77_8_1244752959441 Run Identier York University_79_8_1244834554642 NumRel RelRet 2362 1615 2362 1570 2362 1754 2362 1629 2362 1620 2362 1742 0.2732 0.3516 0.3586 0.2447 0.29 0.2097 MAP 0.3272 0.3375 0.3682 0.3544 truth data of imageCLEF 2008 used to train re-ranking system. Second method we present is an could modied to adopt itself to dataset changes. As a result, we obtain an adaptive re-ranking of information retrieval system if modeled dataset is not subject to change. But proposed technique and evaluate two new methods: rst method is a re-ranking method which considers several aspects of both document and query such as degree of query/document generality, document/query length sum, it is promising that usage of visual term gives better results than most of the textual only etc. System learns to re-rank of initially retrieved documents, using all such features. Ground models. system. On the other hand, experiments on our second proposal that integrates both textual and visual content ranked rst among all submissions for mixed automatic runs.
In this study we proposed an integrated model that ultimately aims to close semantic gap in Results of ImageCLEFmed 2009 shows that proposed re-ranking approach boosts performance integrated retrieval system which extends textual document vectors with visual terms. visual information retrieval. We used a single visual term, however results are satisfactory. In In this paper, we summarize our participation to imageCLEFmed task. In this study, we propose 107E217.
This work is supported by Turkish National Science Foundation (TBT Acknowledgements AK) under project number References 4 Discussion and Future Work Figure 2: Precision-Recall Graph of Top-5 Automatic Mixed Runs of ImageCLEFmed 2009. [1] Erica Chisholm and Tamara G. Kolda. New term weighting formulas for the vector space method in information retrieval. Technical report, 1999.
SIGIR ’96: Proceedings of the 19th annual international ACM SIGIR conference on Research and development in information retrieval, pages 2129, New York, NY, USA, 1996. ACM. [4] Amit Singhal, Chris Buckley, and Mandar Mitra. Pivoted document length normalization. In
In Proceedings of Learning to Rank for Information Retrieval Workshop, pages 5258. ACM, [2] Tao Qin, Tie-Yan Liu, Jun Xu, and Hang Li. How to make letor more useful and reliable.
ACM, 2008.