1. K. Zagoris, K. Ergina and N. Papamarkos Image Processing and Multimedia Laboratory Department of Electrical & Computer Engineering Democritus University of Thrace, 67100 Xanthi, Greece

2. There is a huge information stored in the form of document images These documents do not have any indexing information How we can retrieve a such document using a quary a single word contained in the document 2 Democritus Universiy of Thrace, Greece

3. 3

4. Binarization (Otsu Technique) Original Document Median Filter 4 Democritus Universiy of Thrace, Greece

5. Indentify all the Connected Components (CCs) Calculate the most common height of the document CCs (CC ch ) Reject the CCs with height less than 70% of the CC ch. That only reject areas of punctuation points and noise. Expand the left and right sides of the resulted CCs by 20% of the CC ch The words are the merged overlapping CCs Using the Connected Components Labeling and Filtering method 5 Democritus Universiy of Thrace, Greece

6. Width to Height Ratio Word Area Density. The percentage of the black pixels included in the word-bounding box Center of Gravity. The Euclidean distance from the word’s center of gravity to the upper left corner of the bounding box: 6 Democritus Universiy of Thrace, Greece

7. Vertical Projection. The first twenty (20) coefficients of the Discrete Cosine Transform (DCT) of the smoothed and normalized vertical projection. Original Image The Vertical Projection Smoothed and normalized 7 Democritus Universiy of Thrace, Greece

8. Top – Bottom Shape Projections. A vector of 50 elements The first 25 values are the first 25 coefficients of the smoothed and normalized Top Shape Projection DCT The rest 25 values are equal to the first 25 coefficients of the smoothed and normalized Bottom Shape Projection DCT. 8 Democritus Universiy of Thrace, Greece

9. Upper Grid Features is a ten element vector with binary values which are extracted from the upper part of each word image. Down Grid Features is a ten element vector with binary values which are extracted from the lower part of the word image. 9 Democritus Universiy of Thrace, Greece

10. [0,0,0,1195,0,0,0,0,0,0] [0,0,0,1,0,0,0,0,0,0] [0,0,0,0,0,0,0, 598, 50, 33 ] [0,0,0,0,0,0,0,1,1,0] 10 Democritus Universiy of Thrace, Greece

11. The Structure of the Descriptor 11 Democritus Universiy of Thrace, Greece

12. User enters a query word The proposed system creates an image of the query word with font height equal to the average height of all the word-boxes obtained through the Word Segmentation stage of the Offline operation. For our experimental set the average height is 50 The font type of the query image is Arial The smoothing and normalizing of the various features described before, suppress small differences between various types of fonts 12 Democritus Universiy of Thrace, Greece

13. 13 Democritus Universiy of Thrace, Greece

14. 100 image documents created artificially from various texts Then Gaussian and “Salt and Pepper” noise was added Implement in parallel a text search engine which makes easier the verification and evaluation of the search results of the proposed system 14 Democritus Universiy of Thrace, Greece

15. o Visual Studio 2008 o Microsoft.NET Framework 2.0 o C# Language o Microsoft SQL Server 2005 http://orpheus.ee.duth.gr/irs2_5/ 15 Democritus Universiy of Thrace, Greece

16. o Precision and the Recall metrics o 30 searches in 100 document images o Font Query: Arial  Mean Precision: 87.8%  Mean Recall: 99.26% 16 Democritus Universiy of Thrace, Greece

17. FineReader® 9.0 OCR ProgramQuery Font Name “Tahoma”.  Mean Precision: 76.67%  Mean Recall: 58.42%  Mean Precision: 89.44%  Mean Recall: 88.05% 17 Democritus Universiy of Thrace, Greece

18. The query word is given in text and then transformed to word image The proposed system extract nine (9) powerful features for the description of the word images These features describe satisfactorily the shape of the words while at the same moment they suppress small differences due to noise, size and type of fonts Based on our experiments the proposed system performs better in the same database than a commercial OCR package 18 Democritus Universiy of Thrace, Greece

19. 19 Democritus Universiy of Thrace, Greece