1. Document Image Retrieval LBSC 796/CMSC 828o Douglas W. Oard April 12, 2004 mostly adapted from A lecture by David Doermann

2. Agenda n Questions n Definitions - Document, Image, Retrieval n Document Image Analysis –Page decomposition –Optical character recognition n Traditional Indexing with Conversion –Confusion matrix –Shape codes n Doing things Without Conversion –Duplicate Detection, Classification, Summarization, Abstracting –Keyword spotting, etc n Example: Chinese document images

3. Goals of this Class n Expand your definition of what is a “DOCUMENT” n To get an appreciation of the issues in document image indexing n To look at different ways of solving the same problems with different media n Your job: compare/contrast with other media

4. Document DOCUMENT n Basic Medium for Recording Information n Transient –Space –Time n Multiple Forms –Hardcopy (paper, stone,..) / Electronic (CDROM, Internet, …) –Written/Auditory/Visual (symbolic, scenic) n Access Requirements –Search –Browse –“Read”

5. Sources of Document Images n The Web –Some PDF files come from scanned documents –Arabic news stories are often GIF images n Digital copiers –Produce “corporate memory” as a byproduct n Digitization projects –Provide improved access to hardcopy documents

6. Some Definitions n Modality –A means of expression n Linguistic modalities –Electronic text, printed, handwritten, spoken, signed n Nonlinguistic modalities –Music, drawings, paintings, photographs, video n Media –The means by which the expression reaches you Internet, videotape, paper, canvas, …Internet, videotape, paper, canvas, …

7. Document Images n A collection of dots called “pixels” –Arranged in a grid and called a “bitmap” n Pixels often binary-valued (black, white) –But greyscale or color is sometimes needed n 300 dots per inch (dpi) gives the best results –But images are quite large (1 MB per page) –Faxes are normally 72 dpi n Usually stored in TIFF or PDF format

8. Images n Pixel representation of intensity map n No explicit “content”, only relations n Image analysis –Attempts to mimic human visual behavior –Draw conclusions, hypothesize and verify IMAGE Image databases Use primitive image analysis to represent content Transform semantic queries into “image features” color, shape, texture … spatial relations

9. Document Images n Scanned Pixel representation of document n Data Intensive (100-300dpi, 1-24 bpp) n NO EXPLICIT CONTENT n Document image analysis or manual annotation required –takes pixels -> contents –automatic means are not guaranteed Yet we want to be able to process them like text files! DOCUMENT IMAGE

10. Document Image Database n Collection of scanned images n Need to be available for indexing and retrieval, abstracting, routing, editing, dissemination, interpretation … DOCUMENT DATABASE IMAGE

11. Information Retrieval Document Understanding Document Image Retrieval

12. Managing Document Image Databases n Document Image Databases are often influenced by traditional DB indexing and retrieval philosophies –We are comfortable with them –They work n Problem: Requires content to be accessible n Techniques: –Content based retrieval (keywords, natural language) –Query by structure (logical/physical) –Query by Functional attributes (titles, bold, …) n Requirements: –Ability to Browse, search and read

13. Indexing Page Images (Traditional) Optical Character Recognition Page Decomposition Scanner Document Page Image Structure Representation Character or Shape Codes Text Regions

14. Document Image Analysis n General Flow: –Obtain Image - Digitize –Preprocessing –Feature Extraction –Classification n General Tasks –Logical and Physical Page Structure Analysis –Zone Classification –Language ID –Zone Specific Processing RecognitionRecognition VectorizationVectorization

15. Page Analysis n Skew correction –Based on finding the primary orientation of lines n Image and text region detection –Based on texture and dominant orientation n Structural classification –Infer logical structure from physical layout n Text region classification –Title, author, letterhead, signature block, etc.

16. Image Detection

17. Text Region Detection

18. Language Identification n Language-independent skew detection –Accommodate horizontal and vertical writing n Script class recognition –Asian script have blocky characters –Connected scripts can’t be segmented easily n Language identification –Shape statistics work well for western languages –Competing classifiers work for Asian languages

19. Optical Character Recognition n Pattern-matching approach –Standard approach in commercial systems –Segment individual characters –Recognize using a neural network classifier n Hidden Markov model approach –Experimental approach developed at BBN –Segment into sub-character slices –Limited lookahead to find best character choice –Useful for connected scripts (e.g., Arabic)

20. OCR Accuracy Problems n Character segmentation errors –In English, segmentation often changes “m” to “rn” n Character confusion –Characters with similar shapes often confounded n OCR on copies is much worse than on originals –Pixel bloom, character splitting, binding bend n Uncommon fonts can cause problems –If not used to train a neural network

21. Measures of OCR Accuracy n Character accuracy n Word accuracy n IDF coverage n Query coverage

22. Improving OCR Accuracy n Image preprocessing –Mathematical morphology for bloom and splitting –Particularly important for degraded images n “Voting” between several OCR engines helps –Individual systems depend on specific training data n Linguistic analysis can correct some errors –Use confusion statistics, word lists, syntax, … –But more harmful errors might be introduced

23. OCR Speed n Neural networks take about 10 seconds a page –Hidden Markov models are slower n Voting can improve accuracy –But at a substantial speed penalty n Easy to speed things up with several machines –For example, by batch processing - using desktop computers at night

24. Problem: Logical Page Analysis (Reading Order) n Can be hard to guess in some cases –Newspaper columns, figure captions, appendices, … n Sometimes there are explicit guides –“Continued on page 4” (but page 4 may be big!) n Structural cues can help –Column 1 might continue to column 2 n Content analysis is also useful –Word co-occurrence statistics, syntax analysis

25. Processing Converted Text Typical Document Image Indexing n Convert hardcopy to an “electronic” document –OCR –Page Layout Analysis –Graphics Recognition n Use structure to add metadata n Manually supplement with keywords Use traditional text indexing and retrieval techniques?

26. Information Retrieval on OCR n Requires robust ways of indexing n Statistical methods with large documents work best n Key Evaluations –Success for high quality OCR (Croft et al 1994, Taghva 1994) –Limited success for poor quality OCR (1996 TREC, UNLV) –Clustering successful for > 85% accuracy (Tsuda et al, 1995)

27. Proposed Solutions n Improve OCR n Automatic Correction –Taghva et al, 1994 n Enhance IR techniques –Lopresti and Zhou, 1996 4NGrams Applications –Cornell CS TR Collection (Lagoze et al, 1995) –Degraded Text Simulator (Doermann and Yao, 1995)

28. N-Grams n Powerful, Inexpensive statistical method for characterizing populations n Approach –Split up document into n-character pairs fails –Use traditional indexing representations to perform analysis –“DOCUMENT” -> DOC, OCU, CUM, UME, MEN, ENT n Advantages –Statistically robust to small numbers of errors –Rapid indexing and retrieval –Works from 70%-85% character accuracy where traditional IR fails

29. Matching with OCR Errors n Above 80% character accuracy, use words –With linguistic correction n Between 75% and 80%, use n-grams –With n somewhat shorter than usual –And perhaps with character confusion statistics n Below 75%, use word-length shape codes

30. Handwriting Recognition n With stroke information, can be automated –Basis for input pads n Simple things can be read without strokes –Postal addresses, filled-in forms n Free text requires human interpretation –But repeated recognition is then possible

31. Conversion? n Full Conversion often required n Conversion is difficult! –Noisy data –Complex Layouts –Non-text components Points to Ponder n Do we really need to convert? n Can we expect to fully describe documents without assumptions?

32. Researchers are seeing a progression from full conversion to image based approach n Applications –Indexing and Retrieval –Information Extraction –Duplicate Detection –Clustering (Document Similarity) –Summarization n Advantages –Makes use of powerful image properties (Function, IVC 1998) –Can be cheaper then conversion –Makes use of redundancy in the language.

33. Outline

34. Processing Images of Text n Characteristics –Does not require expensive OCR/Conversion –Applicable to filtering applications –May be more robust to noise n Possible Disadvantages –Application domain may be very limited –Processing time may be an issue if indexing is otherwise required

35. Proper Noun Detection (DeSilva and Hull, 1994) n Problem: Filter proper nouns in images of text –People, Places, Things n Advantages of the Image Domain: –Saves converting all of the text –Allows application of word recognition approaches –Limits post-processing to a subset of words –Able to use features which are not available in the text n Approach: –Identify Word Features Capitalization, location, length, and syntactic categoriesCapitalization, location, length, and syntactic categories –Classify using rule-set –Achieve 75-85% accuracy without conversion

36. Keyword Spotting Keyword Spotting Techniques: –Work Shape/HMM - (Chen et al, 1995) –Word Image Matching - (Trenkle and Vogt, 1993; Hull et al) –Character Stroke Features - (Decurtins and Chen, 1995) 4Shape Coding - (Tanaka and Torii; Spitz 1995; Kia, 1996) Applications: –Filing System (Spitz - SPAM, 1996) –Numerous IR –Processing handwritten documents Formal Evaluation : –Scribble vs. OCR (DeCurtins, SDIUT 1997)

37. Shape Coding n Approach –Use of Generic Character Descriptors –Make Use of Power of Language to resolve ambiguity –Map Character based on Shape features including ascenders, descenders, punctuation and character with holes

38. Shape Codes n Group all characters that have similar shapes –{A, B, C, D, E, F, G, H, I, J, K, L, M, N, O, P, Q, R, S, T, U, V, W, X, Y, Z, 2, 3, 4, 5, 6, 7, 8, 9, 0} –{a, c, e, n, o, r, s, u, v, x, z} –{b, d, h, k, } –{f, t} –{g, p, q, y} –{i, j, l, 1} –{m, w}

39. Why Use Shape Codes? n Can recognize shapes faster than characters –Seconds per page, and very accurate n Preserves recall, but with lower precision –Useful as a first pass in any system n Easily extracted from JPEG-2 images –Because JPEG-2 uses object-based compression

40. Additional Applications n Handwritten Archival Manuscripts –(Manmatha, 1997) n Page Classification –(Decurtins and Chen, 1995) n Matching Handwritten Records –(Ganzberger et al, 1994) n Headline Extraction n Document Image Compression (UMD, 1996-1998)

41. Outline

42. Document Functionality [ICDAR 1997] Humans process documents very robustly n When interacting with documents, we can interpret without recognition n We can judge relevance without reading n We can rapidly navigate documents to find the information we want Claims n We must provide basic ways to interact with documents, and interaction often relies as much on the structure of a document, as on the content n Traditional geometric properties and type-dependent logical models are not sufficient

43. The Role of Documents The role or “function” of a document is to store data in symbolic form which has been produced by a sender (the author) to facilitate transfer to a receiver (the reader) n Documents are designed to be interpreted by humans n Authors typically tailor this design to optimize the transfer of information n Readers use structure to enhance interpretation In what ways does the design facilitate, disambiguate or enhance the flow of information?

45. Functional Structures

48. Outline

49. Graphics n Maps and Drawings –Lorenz and Monagan, 1995 4Samet and Soffer, 1995 –Amlani and Kasturi, 1988 n Graphs – Koga et al, 1993 n Logos and Icons –Jaisimha et al, 1996 –Doermann et al, 1996 –Gudivada and Raghavan, 1993 n Technical Drawings – Syeda-Mahmood, 1995

50. Map Interpretation Samet et al n Identify Legend on the Map Image n Extract Images map labels and descriptions n Identify labels in the map images n Allow user to query based on extracted images n Bootstraps the information extraction and interpretation problems

51. Outline

52. Duplicate Detection n Same content, same format –For example, a xerox copy n Same content, different format –For example, as a web page or on paper n Shared content, same format –For example, a paper with annotations n Shared content, different format –For example, including text with cut-and-paste

53. Duplicate Reconciliation

54. Approach n Use global features to restrict search –Number of pages, number of lines, page moments n Extract a signature – using shape codes n Convert signature –use a set of n-gram keys to index the database n Rank and verify –return top N documents –visual or algorithmic refinement n Advantages: –Robust to noise, extracted quickly, extracted easily, efficiently stored

55. Cross-Language Duplicate Detection (= finding translations!)

56. Evaluation n The usual approach: Model-based evaluation –Apply confusion statistics to an existing collection n A bit better: Print-scan evaluation –Scanning is slow, but availability is no problem n Best: Scan-only evaluation –No existing IR collections have printed materials

57. Summary n Many applications benefit from image based indexing –Less discriminatory features –Features may therefore be easier to compute –More robust to noise –Often computationally more efficient n Many classical IR techniques have application for DIR n Structure as well as content are important for indexing n Preservation of structure is essential for in-depth understanding

58. Example Title Pages (#4 & #9)

59. Title Page Overall Accuracy n 57 Title pages, 891 non-title pages n Overall Accuracy = 906/948 = 95.57% n Title Page Accuracy = 37/57 = 64.91% n False Positives = 22 n False Negatives = 20 n Observations –All without Type-Specific Information –Need Functional (or Logical) Features

60. Agenda n Questions n Definitions - Document, Image, Retrieval n Document Image Analysis n Traditional Indexing with Conversion n Doing things Without Conversion n Recent work on IR with Chinese document images –Tseng and Oard

61. Document Retrieval Approaches for Images of Text n Full-text search based on manually re-keying the text –Prohibitively expensive at large scale n Search based on bibliographic metadata –May be difficult to adequately describe the materials. n Full text based on Optical Character Recognition (OCR) –Inexpensive and relatively rapid –Sensitive to OCR accurracy

62. Key Questions for Information Retrieval n What to index? –Phrase, words, character, or shape codes –Unigrams or n-grams n How to weight a term in a document? –Term frequency (TF) –Document frequency (DF) –Document length normalization –(Term position) n How to assign scores to documents? –Boolean, vector space, and probabilistic models

63. Chinese Text Retrieval Issues n Words may be any number of characters (typically 2-5) –But some that contain only 1 character or more than 5 characters –e.g., “ 貓 ” (cat), “ 聯合國教科文組織 ” (UNESCO) n Longer words (over 2 characters) often have shorter sub-word units –Transliteration is an exception n Written Chinese has no word separator –A sentence can be segmented in different ways, all may be legal –Similar to the phrase detection problem in English n Chinese character inventory is very large –13,500 characters in Big-5 code (traditional Chinese: Taiwan and Hong Kong) –Over 6,000 characters in GB code (simplified Chinese: China, Singapore) –About 3,000 commonly used characters in each character set

64. Socio-Cultural Research Center (SCRC) Collection n 800,000 newspaper clippings from 1950-1976 –Scanned over 300,000 at 300 dpi n 30 China, Hong Kong, and Taiwan news agencies –Mostly simplified Chinese, some traditional Chinese n Focus on diplomatic and military activities

65. Document Preparation n Selected 11,108 scanned document images n OCR yielded 8,438 valid docs (Presto! OCR Pro, Big-5) –Avg valid document had a 69% system-reported “recognition rate” Computed on a sample of 1,300 documentsComputed on a sample of 1,300 documents n Second version prepared using Big-5 to GB conversion –GB version used in experiments

66. Topic Preparation n Based on contemporaneous Chinese journal articles –From 100 paper titles, 30 were selected and rewritten as Chinese topics n Made English translations for cross-language experiments –Translated by native speakers of Chinese 12 Anti-Chinese Movements Activities related to the anti-Chinese movements in Indonesia Articles must deal with activities related to the anti-Chinese movement in Indonesia; case reports or articles dealing with PRC's criticism of the Anti-Chinese movement will be considered partly relevant.

67. Relevance Judgments n Exhaustive tri-state relevance judgments –Irrelevant (=0), partially relevant (=1), fully relevant (=2) n Every topic-document pair judged by 3 assessors –2 majored in history, 1 majored in library science –Averaged 4 minutes per document image (for all 30 topics) n Sum of the judgments provides a final estimate –0=not relevant, 1…5=partially relevant, 6=fully relevant –Threshold as desired to reflect the intended application In our experiments, any score > 0 is treated as “relevant”In our experiments, any score > 0 is treated as “relevant”

68. Chinese OCR Text Retrieval Strategies n Indexing method: –Both 1-gram (for partial match) and 2-gram (for preserving sequence) –Example: “ABC” will be indexed with “A”, “B”, “C”, “AB”, “BC” –Compared to 1-gram only and 2-gram only n Weighting scheme: –document terms : TF*IDF = log(1+ tf ) * log(N/df) –query terms : tf * (3w-1), where w is the length of the term n Retrieval model: –Vector space model compared with probabilistic model n Document length normalization: –byte size for document terms, compared to cosine

69. OCR and Length Normalization n Experiments by Taghva et al showed that –some sophisticated weighting schemes shown to be more effective for ordinary text might lead to more unstable results for OCR degraded text. n Singhal, Salton, Buckley [‘96] analyzed this phenomenon by –Vector space model (SMART system) –Word-based indexing –simulated OCR output of a TREC collection (2GB of 742,202 docs) –50 TREC queries (numbered from 151 to 200) –Specifically, effects of cosine normalization and IDF are analyzed –Incorrect terms like ‘systom’ have large IDF and thus affect weights of other terms in the same document if cosine normalization is used: –They correct this problem by using byte size normalization: (byte size) 0.375

70. Results Summary Long QueriesTitle Queries 1+2 gram is best ByteSize beats Cosine Long queries beat Titles Inquery does well Mean Average Precision

71. Conclusions of Study n The SCRC test collection is useful –But more than 30 topics may be needed for statistical significance n Indexing 1-grams and 2-grams together works well –If 2-grams are given greater weight in the query n Byte size normalization outperforms cosine normalization –But Inquery does better than either on short queries n OCR errors adversely affect blind relevance feedback –A clean comparable collection would probably work better –Pruning seems to help –Considerable parameter tuning is needed ( , , and k)