14
IMPACT Research Image Enhancement, Segmentation, Experimental OCR Apostolos Antonacopoulos PRImA Lab, The University of Salford, United Kingdom www.primaresearch.org

IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Embed Size (px)

Citation preview

Page 1: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

IMPACT ResearchImage Enhancement,Segmentation,Experimental OCR

Apostolos Antonacopoulos

PRImA Lab, The University of Salford, United Kingdom

www.primaresearch.org

Page 2: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Outline Overview: digitisation workflow Image enhancement

Border removal Page curl removal Correction of arbitrary warping

Segmentation Recognition-based Standalone

Typewritten document OCR Wordspotting

2

Page 3: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Overview: Digitisation Workflow

3

Main steps:① Scanning② Image enhancement

Page splitting Border removal Page curl removal Dewarping

③ Layout analysis Segmentation of regions, lines, words and

characters Region classification Logical layout analysis

④ OCR (incl. specialist or wordspotting)⑤ Post-processing

Page 4: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Correction of Arbitrary Warping Fully-automated tool for large-scale

digitisation Interface for interactive fine correction

(e.g. for boutique digitisation projects) Arbitrary geometric artefacts correction Multi-column documents Fully-parameterised process (reversible) No adverse effects on non-warped

documents

22 March 2011 – EC review4

Page 5: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Fully-Automated Dewarping

22 March 2011 – EC review5

Page 6: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Global Grid Construction

22 March 2011 – EC review6

Original Image Region Segmentation Global Grid

Page 7: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Sub-Grid Correction

22 March 2011 – EC review7

Sub-grid text lines

Sub-grid aligned to baselines

Corrected sub-grid

Page 8: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Multi-Column Document Correction

Original image Baseline-aligned sub-grids Corrected image

Page 9: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Preliminary Results

Evaluation calculates deviation from straight lines (shaded area)

Method compared with IMPACT page-curl removal method and with original image

22 March 2011 – EC review9

Page 10: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Textline and Word Segmentation

Standalone methods that can be integrated to systems without the need to integrate FR engine

Not based on recognition of characters/words – suitable for documents with non-dictionary words or not practical to OCR to OCR (word spotting)

Used in other IMPACT methods: Typewritten OCR Correction of arbitrary warping Word spotting

date footertext10

Page 11: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Hybrid Text Line Segmenter Hybrid approach based on connected component clustering and

projection profiles

Connected component extraction (incl. noise filtering)

Group components into line candidates using an efficient data structure

Find and split under-segmented lines using local projection profiles

Merge small peripheral lines to appropriate neighbour (e.g. for i-dots etc.)

Bitonal image

Text regions (PAGE XML)

Regions with text lines (PAGE XML)

Parameters

Page 12: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Density Word Segmenter Adaptive projection-profile based approach using foreground pixel

density

Bitonal image

Text regions and lines (PAGE XML)

Regions, text lines and words (PAGE XML)

Parameters

For each text line: Generate vertical

projection profile Find delimiting white

spaces using an adaptive threshold based on the density of foreground pixels in the line

Group connected components into words

Page 13: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

13

Evaluation Text line ground truth: 25 historical documents (more than 2700 text lines) Results (using USAL layout evaluation tool):

Word ground truth: 15 historical documents (more than 14500 words) Results (using USAL layout evaluation tool):

Page 14: IMPACT Final Conference - Research Parallel Sessions - 01 impact conference_research_session_aa

Further Information14

PRImAhttp://www.primaresearch.org

IMPACThttp://www.impact-project.eu