Text Line Recognition (Course: OCRopus)

Text Line Recognition

Text Line Recognition in OCRopus

primary interface for "OCR": IRecognizeLine
standard mode of operation: takes a "text line" as input, outputs a recognition lattice

text line can be any image with linear arrangement of symbols
direction/geometry of the line image is script dependent
lattice represented in terms of weighted finite state transducers

Text Line Recognition Interface

    struct IRecognizeLine : IComponent {
    virtual void recognizeLine(IGenericFst &result,bytearray &image);
virtual void recognizeLine(intarray &segmentation,IGenericFst &result,bytearray &image);
        ...
    };

interface is much simpler than for ICharacterClassifier

that's because all the results can get stuffed into the result data structure

two versions

first version just performs recognition and outputs a lattice
second version outputs the internally generated oversegmentation
in the second version, the transducer transduces from segmentation components to output characters

Text Line Training Interface

    struct IRecognizeLine : IComponent {
virtual void startTraining(const char *type="adaptation")
virtual void addTrainingLine(bytearray &image,nustring &transcription)
virtual void addTrainingLine(intarray &segmentation, bytearray &image_grayscale, nustring &transcription)
      virtual void finishTraining()
        ...
    };

all training must be surrounded by startTraining ... finishTraining
in startTraining, you indicate whether this is initial training or adaptation
you can train either...

on an unsegmented line image, by giving the line image and the transcription
on a segmented line image, by giving the segmentation, line image, and transcription
a recognizer doesn't have to implement both methods (but it's good if it does)

Text Line Alignment

    struct IRecognizeLine : IComponent {
      virtual void align(nustring &chars,intarray &result,floatarray &costs, bytearray &image,IGenericFst &transcription)
        ...
    };

text line recognizers should also implement an alignment method
this method should output an aligned result using the transcription as the "ground truth"
this is useful for generated isolated character training data

Loading and Saving

    struct ICharacterClassifier : IComponent {
        ...
virtual void save(FILE *stream)
virtual void load(FILE *stream)
    };

as before, you can load and save models

Simple Example

image = read_image_gray_checked(arg[1])
segmenter = make_SegmentPageByRAST()
segmenter:segment(segmentation, image)
recognizer = ocropus_make_RecognizeLine('bpnet', 'models/neural-net-file.nn')
regions = RegionExtractor()
regions:setPageLines(segmentation)
line_image = bytearray()
for i = 1, regions:length() - 1 do
    regions:extract(line_image, image, i, 1)
    fst = make_StandardFst()
    recognizer:recognizeLine(fst, line_image)
    s = nustring()
    fst:bestpath(s)
    print(s:utf8())
end

Course: OCRopus