What are the main reasons for the lack of variety in scene text images?

Scene text images captured by portable cameras feature an abundance of variety due to noise, unusual fonts/typesetting, low resolution, and other factors that make the recognition problem challenging.

What is the advantage of a bigram model?

Such a model has the advantage of being contextual—helping to promote internal consistency among the characters in a parse, while also retaining locality so that efficient dynamic programming algorithms may be used.

What is the parse score for a segment of width w?

Because the dynamic programming algorithm assumes a spatially one-dimensional array of scores, the authors collapse 2-D detected text regions into a line by retaining only the maximum score for each hypothesis (e.g. character label and width) over the rows of each column marked as text.

What is the purpose of this paper?

In this paper, the authors propose the use of additional font-based features in a robust probabilistic model that combines word and character segmentation with recognition.

What is the advantage of using a font-based character bigram model?

This has the advantage of retaining the ability to use an efficient dynamic programming algorithm for inference, while incorporating font properties in the recognition process without the need for large amounts of observed query data.

What is the purpose of the feature?

In addition to the usual character appearance and bigram features, the authors introduce an adaptive, contextual font property feature that does not require prior segmentations or large amounts of query data.

what is the parse score for a segment of width w?

The parse score is the sum of the new discriminants it induces,P (n,w′, w, t, y′, y) = UOn,r + U G n,r + U A r,t (y) W A r,t+ [ UB (y′, y) + UW (y′, y, w′, w) ] WBm,t,(3) where m is the start index of the previous character segment.

Open AccessProceedings Article10.1109/ICPR.2010.970

Typographical Features for Scene Text Recognition

Jerod Weinman

- 23 Aug 2010

- pp 3987-3990

TL;DR: A semi-Markov modelintegrating character segmentation and recognition with a bigram model of character widths is augmented to improve recognition on low-resolution images of signs containing text in many fonts.

Abstract: Scene text images feature an abundance of font style variety but a dearth of data in any given query. Recognition methods must be robust to this variety or adapt to the query data's characteristics. To achieve this, we augment a semi-Markov model---integrating character segmentation and recognition---with a bigram model of character widths. Softly promoting segmentations that exhibit font metrics consistent with those learned from examples, we use the limited information available while avoiding error-prone direct estimates and hard constraints. Incorporating character width bigrams in this fashion improves recognition on low-resolution images of signs containing text in many fonts.

AI Agents for this Paper

Find similar papers on Google Scholar, PubMed and Arxiv
Write a critical review of this paper
Analyze citations of this paper to find unaddressed research gaps

Most frequently asked questions

1. What are the contributions in "Typographical features for scene text recognition" ?

Weinman et al. this paper proposed a semi-Markov model to integrate character segmentation and recognition with a bigram model of character widths.

2. How many fonts are used in the bigram model?

The authors scaled 1600 base training fonts with horizontal aspect ratios scaled at 2− 1 2 , 2− 1 4 , 20, and 2 12 to train UA with 6400 fonts.

3. How many character widths are trained in the bigram model?

The authors quantized character widths to w ∈ {4, 8, 12, 16, 20, 24, 32} pixels and trained the width bigram model UW using the frequency of association of between pairs of character labels and quantized widths in the 1866 training fonts.

4. What is the significance of the character widths?

If these character widths exhibit regular patterns, they should be taken advantage of during the segmentation/recognition process.

Figure 2. TOP: Scatter plots of character widths over several fonts for the most correlated ( 6 and 9, left) and more typically correlated ( A and m, right) character pairs. A fixed font size is used. BOTTOM: Histogram of correlation coefficients for widths of all characters pairs.

Figure 3. Examples of sign images correctly recognized by our model (top) and those remaining a challenge (bottom).

Table 1. Recognition results with increasing amounts of information.

Citations

•Proceedings Article•10.5244/C.26.127

Scene Text Recognition using Higher Order Language Priors

Anand Mishra, +2 more

- 07 Sep 2009

TL;DR: A framework is presented that uses a higher order prior computed from an English dictionary to recognize a word, which may or may not be a part of the dictionary, and achieves significant improvement in word recognition accuracies without using a restricted word list.

...read moreread less

•Journal Article•10.1109/TPAMI.2014.2366765

Text Detection and Recognition in Imagery: A Survey

Qixiang Ye, +1 more

- 01 Jul 2015

- IEEE Transactions on Pattern Analysis an...

TL;DR: This review provides a fundamental comparison and analysis of the remaining problems in the field and summarizes the fundamental problems and enumerates factors that should be considered when addressing these problems.

...read moreread less

876

•10.7275/5474920.0

Improving Text Recognition in Images of Natural Scenes

Jacqueline L Feild

- 01 Jan 2014

TL;DR: This work aims to contribute to the understanding of language recognition in the context of natural settings by improving text recognition in images of natural scenes.

...read moreread less

Patent

Name search using multiple bitmap distributions

David Edward Biesenbach, +3 more

- 12 Apr 2012

TL;DR: In this article, it is determined whether a first bitmap signature of a query name and a second bitmap signatures of a target name have a number of character n-grams overlapping that meet or exceed a threshold to generate a first preliminary value.

...read moreread less

•10.25139/DINAMIKA.V7I1.1568

Listening Comprehension: The impact of Subtitles Typographical Features on Game Visual Novel

Eny Kusumawati

- 10 May 2019

TL;DR: In this article, the impact of typographical features of subtitles including size, colour and position on L2 English viewers' retention and recall of texts in game visual novel, and also the voice effect of the visual game.

...read moreread less

...

Expand

References

•Book

An introduction to digital image processing

Wayne Niblack

- 01 Jan 1986

Proceedings Article•10.1109/ICIP.1995.537667

The steerable pyramid: a flexible architecture for multi-scale derivative computation

Eero P. Simoncelli, +1 more

- 23 Oct 1995

TL;DR: An architecture for efficient and accurate linear decomposition of an image into scale and orientation subbands and the construction and implementation of the transform is described.

...read moreread less

1.3K

•Proceedings Article•10.1109/CVPR.2004.1315187

Detecting and reading text in natural scenes

Xiangrong Chen, +1 more

- 27 Jun 2004

TL;DR: The overall algorithm has a success rate of over 90% (evaluated by complete detection and reading of the text) on the test set and the unread text is typically small and distant from the viewer.

...read moreread less

744

•Proceedings Article

Piecewise training for undirected models

Charles Sutton, +1 more

- 26 Jul 2005

TL;DR: In this article, the authors show that piecewise training is more accurate than pseudolikelihood, and often performs comparably to global training using belief propagation, and they show that this piecewise method can be justified as minimizing a new family of upper bounds on the log partition function.

...read moreread less

191

Journal Article•10.1006/CVIU.1996.0005

Font and Function Word Identification in Document Recognition

Siamak Khoubyari, +1 more

- 01 Jan 1996

- Computer Vision and Image Understanding

TL;DR: An algorithm is presented that identifies the predominant font in which the running text in an English language document is printed, and the repeated words in the document are utilized to overcome noise in the input.

...read moreread less

118

Typographical Features for Scene Text Recognition

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What are the contributions in "Typographical features for scene text recognition" ?

2. How many fonts are used in the bigram model?

3. How many character widths are trained in the bigram model?

4. What is the significance of the character widths?

5. What are the main reasons for the lack of variety in scene text images?

6. What is the advantage of a bigram model?

7. What is the parse score for a segment of width w?

8. What is the purpose of this paper?

9. What is the advantage of using a font-based character bigram model?

10. What is the purpose of the feature?

11. what is the parse score for a segment of width w?

Figures

Citations

Scene Text Recognition using Higher Order Language Priors

Text Detection and Recognition in Imagery: A Survey

Improving Text Recognition in Images of Natural Scenes

Name search using multiple bitmap distributions

Listening Comprehension: The impact of Subtitles Typographical Features on Game Visual Novel

References

An introduction to digital image processing

The steerable pyramid: a flexible architecture for multi-scale derivative computation

Detecting and reading text in natural scenes

Piecewise training for undirected models

Font and Function Word Identification in Document Recognition

Related Papers (5)

A method for text localization and recognition in real-world images

End-to-end scene text recognition

Scene Text Recognition Using Similarity and a Lexicon with Sparse Belief Propagation

Word spotting in the wild

Character recognition in natural images