If you are having trouble with the sign in process, please <ContactLink>contact the SSH Open Marketplace team</ContactLink>.

If you are having trouble with the sign up process, please <ContactLink>contact the SSH Open Marketplace team</ContactLink>.

Subject

Social Sciences & Humanities Open Marketplace

Successfully submitted!

First you need to precise what is the purpose of your text extraction task. There are many possible goals of text extraction, e.g. Get searchable text: plain text ocr, without page segmentation, e.g. in order to use it in search engine or statistical analysis of textual data „reading text" (reconstructing the main text flow excluding running heads etc.) e.g. to use it in further text collation exercises reconstruct page layout with all the details, e.g. to use it for further annotation exercises It is also important to define text representation guidelines, or normalisation choices.

The important characteristics of the image are the following: Layout elements (tables, graphics, ...) Script type (handwritten, Gothic script, Roman script, non-European) Mixed content (pre-printed documents with handwriting) These characteristics determine the method: If it's a clean printed text, a classic OCR engine would be enough In other cases, i.e. handwritten, old printed texts with blackletter (gothic) scripts, or low quality scans, using a model is necessary. If a model exists, or it works out of the box, or it needs to be retrained or adapted. Otherwise, you have to train a model from scratch.

If you choose a classic commercial OCR engine like ABBYY FineReader, the customization would be limited. Others tools allow for more options (training a model for instance). It may be interesting also to use separate tools to perform different steps of the recognition process (Layout and character recognition)...

To identify specific text structures, it may be suitable to use a separate engine, for instance for table recognition, or to detect columns.

In case all existing engine give poor results or if the desired textual representation is very specific, one has to create his/her own training model or adapt an existing one. For this, one needs a number of manually transcribed pages: the larger part is used as the training set for an OCR/HTR engine; a smaller one is used to evaluate the result of the trained model. Such transcriptions can be obtained by several ways: It can be produced internally by the project, it could be outsourced to a contractor, or it could be produced by volunteers. Caveats: Take care of the character encoding It can be good practice to test 2 engines and merge models (but be careful at the overfitting problem). Printed material: No need of a lot of training data (random 5 to 10 pages)

The first step to take is the definition of a target accuracy. Clean printed text: WER 70% is a minimum (can be 60% for a search engine) HTR of badly printed: The acceptable WER may be less. Then again, it depends on the outcomes. To test the accuracy of the model, two metrics are available: Character Error rate (CER) and Word Error rate (WER). WER is more important because it defines the readability of the text. It is also important to assess the quality of the layout recognition.

it may be a good practice to correct a subset of the output, in order to acquire additional training data, when the output is too far from the target accuracy. The post correction can be made either manually or with help of dedicated tools.

Usually, retrain the model with the corrected output from the first recognition iteration is very valuable. The number of pages to correct depends on the input, but the same amount of pages than the initial training data is reasonable.

Differents data formats for OCR outputs. The most common formats are ALTO, hOCR and PAGE. All these formats are expressed in XML, so they are easily convertible to one another.

Convert the output to TEI, a format able to represent the semantics.

The visualization of OCR formats is made possible by various softwares. The choice is function of the desired outcome (searchable PDF, structured XML, etc.)

Extract textual content from images

Media

Workflow steps(13)

1 Define the characteristics of the outcome

2 Define the characteristics of the image

3 Survey existing experiences

4 Choose engine based on the type of the content

5 Layout analysis

6 Create manual transcriptions

7 Training the model

8 Test on a subset and assess quality

9 Correct output

10 Re-train the model with corrected output

11 Produce OCR output in standardized format

12 Extract the structure information from recognized blocks

13 A few visualization options