Abstr act. We present a report about our participation in the ImageCLEF photo task 2006 and a short description of our new framework for future use in further CLEF participations. We described and analysed our participation in the monolingual English task. Special Lucenealigned query expansion and histogram comparisons are helping to improve the baseline results. ACM Categor ies and Subject Descr iptor s 1 Intr oduction Our goal was to develop a new information retrieval framework which can handle different custom retrieval mechanisms (i.e. Lucene, the GNU ImageFinding Tool and others). The framework should be so general that one can use either one of the custom retrieval mechanisms or combine different ones. Therefore our efforts were not focused on better results but to get the framework up and running. 2 System descr iption For our participation at CLEF 2006 we constructed a GUI (graphical user interface) quite easy to handle. It bases on Lucene version 1.4 and includes several userconfigurable components. For example we wrote an Analyzer in which TokenFilter s can be added and rearranged. For the participation in ImageCLEF 2006 we needed to extend this system with image retrieval capabilities. But instead of extending the old system we rewrote it and developed a general IR framework. This framework does not necessarily require Lucene and makes it possible to build an abstract layer for other search engines like GIFT (GNU ImageFinding Tool) and others. The main component of our framework is the class Run, which includes all information required to perform and repeat searches. A Run contains the Topic s, which can be preprocessed by TopicFilters a search method called Searcher , which can read and search the Index and HitSet s, which contain the results. To use another search engine one will have to extend the abstract classes Indexer and Searcher . To use another data collection type (e.g. GIRT4, IAPR, ...) one will have to extend the abstract class DataCollection.
an Index, which was created by an Indexer from a DataCollection In order to merge results there is a special extension of the class Searcher : the MergeSearcher . It combines several Searcher s and uses an abstract class Merger to merge their results.
The text search engine used is Lucene version 1.9 with an adapted Analyzer . As described above we create a LuceneIndexer and an abstract LuceneSearcher that provides basic abilities to underlying Searcher s like a selectable Lucene Analyzer . The Analyzer is based on the StandardTokenizer and LowerCaseFilter by Lucene and is enhanced by a custom Snowball filter for stemming. We also added a positional stop word filter and used the stop word list suggested on the CLEF website. The Analyzer details (stemmer and stop word list) can be configured through a GUI. The LuceneStandardSearcher works with the Lucene MultiFieldQueryParser and searches in specified fields of the index.
The GIFT was not used for a submitted run but it should be mentioned here to show the abilities of the framework. We only implemented a Searcher , which can access a GIFT server and takes images to generate queries by example. The underlying index must be created with the command line utility provided by the GIFT.
To access the IAPR data collection of ImageCLEF we implemented the class IaprDataCollection that provides two special options for the DataDocument s to contain special data: · · option 1: the DataDocument s contain the annotations option 2: the DataDocument s contain the histograms In both cases the associated images can be retrieved with the help of the DataDocument .
We focused on the monolingual task using the English topics and annotations. The Lucene index of the annotations contains all fields. All fields are indexed and tokenized except DATE, which is not tokenized. The DOCNO is the only field which is stored, resulting in a quite small index (only 2 MB). We used the default BooleanQuery from Lucene with no further weighting of any field pairs in order to avoid any corpus specific weighting. The default Lucene search classes (default similarity, etc.) were used to search the topicannotation field pairs as shown in table 1.
For all runs, except “tucEEANT”, the query expansion was applied once, using the first 20 results of the original query. In the run “tucEEAFT2” the query expansion is applied twice. The first time using the annotations of the example images and the second time using the first 20 results of the original query. To speed up the colour histogram comparison, we also created an index for the histogram data. For each image
As one can see, run “tucEEAFT2” is our best run which is no surprise if one takes into account that three definitely relevant documents are used for query expansion. Our second best run is “tucEEAFTI” which shows that the additional histogram comparison produces better results than without it. The good results in ImageCLEF 2006 show that our framework works and that our first implemented search methods provide good results. The normal query expansion improves the mean average precision by more than 1%. Considering the poor overall mean average precision this seems to be a good improvement of the results. Although the histogram comparison is a very simple and presumably poor measurement to compare images, it improves our textonly retrieval slightly. The next steps to improve the results is to implement a better image retrieval algorithm due to the weak performance of the histogram comparison.