Abstract
Recent book digitization initiatives have facilitated the access and search of millions of books. Although OCR remains essential for retrieving printed documents, OCR engines remain limited in the languages they handle and are generally expensive to build. This paper proposes a language independent approach that enables search through printed documents in a way that combines image-based matching with conventional IR techniques without using OCR. While image-based matching can be effective in finding similar words, complementing it with efficient retrieval techniques allows for sub-word matching, term weighting, and document ranking. The basic idea is that similar connected elements in printed documents are clustered and represented with ID's, which are then used to generate equivalent textual representations. The resultant representations are indexed using an IR engine and searched using the equivalent ID's of the connected elements in queries. Though, the main benefit of the proposed approach lies in languages for which no OCR exists, the technique was tested on English and Arabic to ascertain the relative effectiveness of the approach. The approach achieves more than 61% relative effectiveness compared to using OCR for both languages. While the reported numbers are lower than that of OCR-based approaches, the proposed method is fully automated, does not require any supervised training, and allows documents to be searchable within a few hours.
Original language | English |
---|---|
Title of host publication | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
Pages | 334-343 |
Number of pages | 10 |
Volume | 5721 LNCS |
DOIs | |
Publication status | Published - 9 Nov 2009 |
Externally published | Yes |
Event | 16th International Symposium on String Processing and Information Retrieval, SPIRE 2009 - Saariselka, Finland Duration: 25 Aug 2009 → 27 Aug 2009 |
Publication series
Name | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
---|---|
Volume | 5721 LNCS |
ISSN (Print) | 03029743 |
ISSN (Electronic) | 16113349 |
Other
Other | 16th International Symposium on String Processing and Information Retrieval, SPIRE 2009 |
---|---|
Country | Finland |
City | Saariselka |
Period | 25/8/09 → 27/8/09 |
Fingerprint
Keywords
- Image based retrieval
- OCR
- Printed documents retrieval
ASJC Scopus subject areas
- Computer Science(all)
- Theoretical Computer Science
Cite this
Efficient language-independent retrieval of printed documents without OCR. / Magdy, Walid; Darwish, Kareem; El-Saban, Motaz.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Vol. 5721 LNCS 2009. p. 334-343 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 5721 LNCS).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution
}
TY - GEN
T1 - Efficient language-independent retrieval of printed documents without OCR
AU - Magdy, Walid
AU - Darwish, Kareem
AU - El-Saban, Motaz
PY - 2009/11/9
Y1 - 2009/11/9
N2 - Recent book digitization initiatives have facilitated the access and search of millions of books. Although OCR remains essential for retrieving printed documents, OCR engines remain limited in the languages they handle and are generally expensive to build. This paper proposes a language independent approach that enables search through printed documents in a way that combines image-based matching with conventional IR techniques without using OCR. While image-based matching can be effective in finding similar words, complementing it with efficient retrieval techniques allows for sub-word matching, term weighting, and document ranking. The basic idea is that similar connected elements in printed documents are clustered and represented with ID's, which are then used to generate equivalent textual representations. The resultant representations are indexed using an IR engine and searched using the equivalent ID's of the connected elements in queries. Though, the main benefit of the proposed approach lies in languages for which no OCR exists, the technique was tested on English and Arabic to ascertain the relative effectiveness of the approach. The approach achieves more than 61% relative effectiveness compared to using OCR for both languages. While the reported numbers are lower than that of OCR-based approaches, the proposed method is fully automated, does not require any supervised training, and allows documents to be searchable within a few hours.
AB - Recent book digitization initiatives have facilitated the access and search of millions of books. Although OCR remains essential for retrieving printed documents, OCR engines remain limited in the languages they handle and are generally expensive to build. This paper proposes a language independent approach that enables search through printed documents in a way that combines image-based matching with conventional IR techniques without using OCR. While image-based matching can be effective in finding similar words, complementing it with efficient retrieval techniques allows for sub-word matching, term weighting, and document ranking. The basic idea is that similar connected elements in printed documents are clustered and represented with ID's, which are then used to generate equivalent textual representations. The resultant representations are indexed using an IR engine and searched using the equivalent ID's of the connected elements in queries. Though, the main benefit of the proposed approach lies in languages for which no OCR exists, the technique was tested on English and Arabic to ascertain the relative effectiveness of the approach. The approach achieves more than 61% relative effectiveness compared to using OCR for both languages. While the reported numbers are lower than that of OCR-based approaches, the proposed method is fully automated, does not require any supervised training, and allows documents to be searchable within a few hours.
KW - Image based retrieval
KW - OCR
KW - Printed documents retrieval
UR - http://www.scopus.com/inward/record.url?scp=70350676786&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=70350676786&partnerID=8YFLogxK
U2 - 10.1007/978-3-642-03784-9_33
DO - 10.1007/978-3-642-03784-9_33
M3 - Conference contribution
AN - SCOPUS:70350676786
SN - 3642037836
SN - 9783642037832
VL - 5721 LNCS
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 334
EP - 343
BT - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
ER -