OCR

This group contains functions for Optical Character Recognition.

The OCR functions identify characters in an image and return the recognized characters as a string. Depending on the model, the classification is based on either classical or convolutional neural network (CNN) machine learning models. The models are provided in a trained form so that the functions can be used directly. No additional settings or training are required. The supported characters and fonts depend on the data used in the training. For this reason, there are various models whose functionalities, as well as the general preconditions and requirements, are described below.

For all characters that are recognized in an image, the models assign each character to a known class with the highest match. The result therefore always consists of the known characters of the model used. By default, rejection or removal of unknown characters is not included. Optionally, in the Expert function for classic machine learning models, you can select a novelty detector that filters out or rejects unknown characters.

The functions only support single-line character strings; several lines must be divided into individual ROIs and read with several function calls.

General requirements for recognizing characters:

• character height min. 20 pixels
• line width min. 3 pixels
• points min. 3 x 3 pixels
• lines min. 3 x 6 pixels
• character spacing / separation min. 4 pixels
• Characters must not overlap
• only horizontal alignment / arrangement of characters max. ± 6° deviation
• The line/contour of a character must not be interrupted
• The characters must not be mirrored or upside down

General requirements for the image:

• ROI only with the characters to be recognized and an interference-free zone around the characters
• Within the ROI, the characters must not be enclosed by another contour such as a rectangle
• good contrast between character and background
• uniform, opaque background free of noise and interference

Requirements for the fonts:

• only proportional fonts with equal character widths
• Larger gaps are always recognized as a single space
• only sans serif fonts such as Arial, Tahoma, Courier, Univers, Frutiger, Verdana, OCR-B
• no mixed fonts
• no dot print or italic fonts

Models:

The enum ETcVnOcrModelType allows access to the following models:

• OMT_NUMBERS
• enables the classification of numbers
• contains the characters of 0-9
• OMT_NUMBERS_SC
• enables the classification of numbers and special characters
• contains the characters from 0-9 and 6 special characters . / - : = +
• OMT_UCLETTERS
• enables the classification of capital letters
• contains the characters of A-Z
• OMT_NUMBERS_SC_UCLETTERS
• enables the classification of numbers, special characters, and capital letters
• contains the characters from 0-9, the 6 special characters . / - : = +, and A-Z

When using the combined model (OMT_NUMBERS_SC_UCLETTERS), confusion can occur due to the great similarity of certain characters. Examples of this are O and 0, S and 5 and B, and 8.

If the position of a number or letter is known, the Expert function with the format specification (sPattern) and the combination of the separate models should be used as an alternative.

• OMT_CNN_NUMBERS_SC_LETTERS
• allows for the classification of numbers, special characters, and uppercase and lowercase letters
• contains the characters from 0-9, the special characters . / - : + ! & $ ( ) %, A-Z , and a-z

When using this model, confusion can occur due to the great similarity of certain characters. Typical examples of this are O and 0, S and s, and X and x.

In general, after a function call, the characters in the result string (ipCharacters) can be analyzed individually and processed afterward, allowing you to decide on a case-by-case basis whether, for example, a 0 is accepted instead of an O.

Initialize functions

In order to be able to use the OCR function with one or more models, the function must first be initialized with the respective models using the function block InitializeFunction.

The time required to load the models depends on the performance of the IPC used and the model size and can take several hundred milliseconds or even several seconds. Therefore, cycle time overruns are to be expected with shorter task cycle times.

If the combined model OMT_NUMBERS_SC_UCLETTERS is to be used, the router memory must be set to at least 512 MB due to the file size.

The exact size depends on the other use of the router memory and any existing fragmentation. If the function block InitializeFunction returns a return code 0x80004005, for example, this indicates insufficient router memory. The router memory should therefore be checked first and increased if necessary. After loading a model, a large part of the memory is released again and is available for the application.

Interpretation of the HRESULT

• The standard function returns S_OK if characters are recognized on the image. When using optional parameters, it depends on whether a sPattern has been specified. If no formatting specification has been defined, the return corresponds to the standard function. If a formatting specification has been passed, S_OK is only returned if the recognized characters match the pattern specification.
• If the standard function was executed successfully and no characters were found or if the characters found do not match the optional pattern specification, S_FALSE is returned.
• If the transferred model is missing for an optional sPattern specification, a HRESULT = ADSERR_DEVICE_NOTINIT is returned.
• If several models are passed to the standard function, a HRESULT = ADSERR_DEVICE_INVALIDPARM is always returned, as the function only supports one model per call. When using optional parameters, it depends on whether a sPattern has been specified. If sPattern is empty, only the TCVN_OMT_NUMBERS_SC_UCLETTERS model is used. If this model has not been passed, a HRESULT = ADSERR_DEVICE_INVALIDPARM is returned.

Examples for the use of sPattern

Additional possibilities arise when using the optional parameters sPattern and eOcrOptions. Depending on the combination of characters on the input image, the information from sPattern, and eOcrOptions, and the recognized characters, S_FALSE is returned if there is a mismatch. The characters of the result string (ipCharacters) can then also be analyzed individually, as S_FALSE is also returned if the length differs, for example. The following examples illustrate the use of the OMT_NUMBERS_SC_UCLETTERS model, which supports numbers, special characters, and uppercase letters. The OMT_CNN_NUMBERS_SC_LETTERS model does not support the use of sPattern.