OCR AI Learn

Applies to version: 2020.1.x; author: Wojciech Kołodziej

Introduction

The OCR AI Learn mechanism allows you to create dedicated templates for recognizing documents in the process. Depending on the type of processed documents, various different distinguishers are used, which clearly indicate which template to use for the document. For example, in the invoice process – the contractor’s Tax ID is the most used. If in the process there are several types of documents, it’s a  good idea to create a distinguisher based on the Tax ID  and acronym of the form type. During the teaching process, the mechanism chooses the areas where values are most likely to occur and keywords that are most likely to be found within the vicinity of the value.

This article describes the “OCR AI learn” process. For more information about OCR, check out the following articles:

• The OCR verification view for the MODERN form
• New OCR verification view
• Custom OCR fields
• OCR AI – Installation

If the teaching process is started for the first time for a given contractor, a new dedicated template is created. The list of all templates can be found in the “System settings” tab in WEBCON BPS Designer Studio.

Fig. 1. OCR AI project in the “System settings” tab

If there was already a template for the contractor, it is trained during teaching and a new version of the template is created.

Fig. 2. OCR projects management

How to create a new document recognition template?

The process of creating or teaching OCR AI templates consists of several stages:

1. Selection of elements to teach

A user validates the recognition of the values. If a value is corrected,  you can select it for the teaching process – just check the field in the “Teach” column.

All selected fields will be used to correct the template for a given contractor.

Fig. 3.Teaching based on the contractor’s invoice

If no field is selected for teaching – the process will no be started.

2. Creating a new template based on existing documents

In the teaching process, a maximum of 100 most recent documents will be used which:

• have been registered earlier than the current document and are or were in the Verification step
• are issued by the same contractor
• have the same document type in the process e.g. cost invoice

If you need a non-standard way of selecting documents for the teaching process, e.g. If you want to  teach based on  2 document types or only documents from the verification step – you can change the attachments selection criteria by using the SQL query.

Fig. 4. The example criteria of attachments selection

Remember to include the distinguisher (if the dedicated network mechanisms are used) and document type in the query – otherwise, all attachments that have passed the OCR AI recognition process will be used.

You can see the created templates in the “OCR AI Learn Queue” report in WEBCON BPS Designer Studio.

Fig. 5. OCR AI learn queue

3. The new template will appear in OCR AI projects in the “System settings” tab

By clicking on the template from the selected provider, you can see which version of the template has been created.

Fig. 6. The example OCR AI project

From the moment a new template appears, it will be used to recognize new documents. If you want to restore to an earlier version:

• Select the fields for correction
• Select the version you want to restore
• Click the “Restore” button to confirm the changes

Important tips

• Scan quality (text layer) – a document scan should be legible and have the correct text layer. The quality of this layer can be verified by checking how the “blocks” around the text are created. If they correctly indicate single values and it is possible to copy them to the fields on the form, the text layer has been created correctly. If the values inside the blocks are stuck together or the blocks are uneven then the document is not suitable for teaching.

Fig. 7. The example scan that is not suitable for teaching

• The document must be entirely electronic. The handwritten documents prevent the correct creation of the text layer.
• The number of example documents – the teaching process is best done when the database already has several correctly verified documents from a given contractor.
• Correct recognition of the searched value – use the teaching mechanism only if the searched value appears in the document and is correctly copied to the field on the form. If individual characters of a given value are incorrect (e.g. 8 instead of B), it should not be used in the teaching process.

Example:

If a suitable place was specified during the recognition, but after copying the characters are wrong - the template should not be taught. Bad characters are the result of poor scan quality or unusual fonts.

Fig. 8. The example invoice

• Thorough verification of previous documents – in the OCR AI Learn process all documents from a given contractor are used (with the same distinguisher) that were registered before the current document. If they have not been verified, or incorrect fields have been indicated, then teaching may deteriorate the quality of field recognition.

Example:

If on the first 10 invoices, the gross amount is mistakenly based on the Taxable Amount. Even if we rectify this in the 11^th invoice and indicate the gross amount from the Invoice Total line, the template will continue to collect the value from the Taxable Amount field.

In the teaching process, 10 incorrect verified invoices will be used and only one correct. Therefore, thorough verification of all documents is crucial. The place of recognition of the value will be corrected when the number of correctly verified documents in the database is significantly greater than the number of incorrect documents.

Fig. 9. The example Invoice

• The correct format of the searched value - the searched value must be in the correct format on the document, otherwise there is no sense in teaching.

For example, if the due date was entered as:

Fig. 10. The example invoice

it will not be possible to copy it to the date field on the form.

Likewise, if the values are incorrectly formatted on the document. In the example below, the date is written together with the year, so both the date and the "r" (Description field) are copied, which prevents this value from being entered into a date field.

Fig. 11. The example invoice

• Stamps / signatures / barcodes - if there is a stamp or a handwritten signature on the document near the values searched, it may prevent the correct recognition of the text in this area.

Fig. 12. The example Invoice

• Data sources/dictionaries - if the data source is used for recognition, check if the value you are looking for is in the source. If a given value is not present, then it should be entered into the data source -there is no point in teaching the document.
• Alphabet / language used in the document - if more than one language / alphabet is used in the document, it may make it difficult to correctly recognize the values.
• Frequency of teaching - all documents from a given contractor are used during teaching - if this process is performed after each new document, you will teach many times based on the same data. So, it is recommended to verify each document, and to initiate the teaching process when there are already several verified documents from a given contractor.