You are here: Start » Function Reference » Computer Vision » Optical Character Recognition » TrainOcr_MLP

TrainOcr_MLP

Header:	FIL.h
Namespace:	fil
Module:	OCR

Trains an OCR multilayer perceptron classifier.

Syntax

void fil::TrainOcr_MLP
(
	const ftl::Array<fil::CharacterSample>& inCharacterSamples,
	const fil::Size& inNormalizationSize,
	ftl::Optional<const ftl::Array<int>& > inHiddenLayerSizes,
	ftl::Optional<int> inRandomSeed,
	const fil::CharacterFeatures& inCharacterFeatures,
	float inLearningRate,
	float inMomentum,
	int inIterationCount,
	ftl::Optional<const fil::Size&> inCharacterSize,
	fil::OcrModel& outOcrModel,
	float& outTrainingAccuracy,
	fil::Profile& diagError,
	ftl::Array<fil::Image>& diagNormalizedCharacters
)

Parameters

Name	Type	Range	Default	Description
inCharacterSamples	const Array<CharacterSample>&			Training font created from sample regions
inNormalizationSize	const Size&		(Width: 16, Height: 16)	The character size after normalization
inHiddenLayerSizes	Optional<const Array<int>& >		NIL	Internal structure of neuron layers used in classifier
inRandomSeed	Optional<int>	0 - +	NIL	Random seed used by MLP classifier
inCharacterFeatures	const CharacterFeatures&		(Pixels: True)	Character features used to distinguish characters from each other
inLearningRate	float	0.01 - 1.0	0.6f	Suppression level of changes during learning process
inMomentum	float	0.0 - 1.0	0.75f	Value of classifier learning momentum
inIterationCount	int	1 - +	100	Learning iteration count
inCharacterSize	Optional<const Size&>		NIL	Size of fixed width font
outOcrModel	OcrModel&			Trained OcrMlpModel used to recognize characters
outTrainingAccuracy	float&			The overall training score
diagError	Profile&			Changes of mean error level progress during learning process
diagNormalizedCharacters	Array<Image>&			Images of normalized characters used to train classifier

Description

This filter prepares a MLP classifier for the further OCR operations.

Filter requires a set of prepared CharacterSample which can be created using MakeCharacterSamples.

Parameter inCharacterSize defines the size of character cropping box. It is especially useful when characters are much bigger than normalization size. When it has Nil value the character is cropped to its bounding box.

Parameters inHiddenLayerSizes, inRandomSeed are used in the process of learning of a newly created MLP classifier. For further parameter description please refer to the documentation of the MLP_Init filter.

All the input regions in filters RecognizeCharacters and TrainOcr_MLP are resized to the size specified in the inNormalizationSize input parameter. The further classification is performed on the normalized regions. Therefore, it is important to select the appropriate normalization size.

The selection of too small normalization size may result in loss of character details. However, too large value of normalization size increases the classifier learning time. The best recognition results are obtained when the size of character is nearly the same as the normalization size.

The character classification depends on character features that are selected in the inCharacterFeatures parameter. At least one feature must be selected. By the default the feature Pixels is selected.

The table below contains the description of each available character feature:

Feature name	Description	Filter origin	Normalized
Pixels	Values of the image pixels after normalization.		False
NormalizedPixels	Values of the image pixels after normalization normalized to range <0, 1.0>.		True
Convexity	Ratio of the input region area to area of its convex hull.	RegionConvexity	True
Circularity	Ratio of the region area to area of its bounding circle.	RegionCircularity	True
NumberOfHoles	Number of holes found in the input region.	RegionHoles	True
AspectRatio	Ratio of input region width to its height.	RegionBoundingBox	False
Width	Region bounding box width.	RegionBoundingBox	False
Height	Region bounding box height.	RegionBoundingBox	False
AreaRatio	Ratio of the input region area to area of its bounding box.		True
DiameterRatio	Ratio of the input region diameter to diameter of its bounding box.	RegionDiameter	True
Elongation	Ratio of longer axis of the approximating ellipse to the shorter one.	RegionElongation	False
Orientation	Further details in the filter RegionOrientation documentation.	RegionOrientation	True
Zoning4x4	Normalized pixel values of region reduced to size 4x4 pixel.		True
HorizontalProjection	Values of normalized image projection normalized by region height.	ImageProjection	True
VerticalProjection	Values of normalized image projection normalized by region height.	ImageProjection	True
HoughCircles	Count of circles found in the normalized image.		True
Moment_11	Character geometric moment type M₁₁.	RegionMoment	False
Moment_20	Character geometric moment type M₂₀.	RegionMoment	False
Moment_02	Character geometric moment type M₀₂.	RegionMoment	False

Remarks

It is recommended not to set normalization size greater than 50 pixels in each dimension. That could make learning time too long.

For more remarks about using MLP classifier please refer to the documentation of the MLP_Init filter.

To read more about how to use traditional OCR technique, refer to Machine Vision Guide: Optical Character Recognition - traditional method

Errors

List of possible exceptions: