OCR Interfaces

• key processing steps in OCRopus are available through simple C++ interfaces
• defined in colib/ocrinterfaces.h
• all you need in order to implement these is colib (a few header files)
• eventually, will permit dynamic loading of components
• eventually, will be able to implement these in Lua
  

OCR Interfaces

These are the standard "pluggable" component interfaces for OCRopus; if you implement one of these, your code should be usable directly in the OCRopus top-level loop:

• IComponent -- base class for all OCR interfaces
  
• ICleanupGray -- improve quality of gray scale images in some way
  
• ICleanupBinary -- improve quality of binary images in some way
  
• ITextImageClassification -- classify regions on a page into text and non-text
  
• IBinarize -- binarize grayscale images
  
• ISegmentPage -- segment a page into columns, lines, headers, footers, etc.
  
• ISegmentLine -- segment a line into character parts
  
• IGenericFst -- generic language model
  
• ICharacterClassifier -- isolated character classification
  
• IRecognizeLine -- text line recognition
  

IComponent

struct IComponent {         virtual const char *description() = 0;         virtual void set(const char *key,const char *value) ...         virtual void set(const char *key,double value) ...         virtual const char *gets(const char *key) ...         virtual double getd(const char *key) ...         virtual ~IComponent() {}     };

method to identify the component at runtime methods to set parameters by name methods to get parameters and statistics by name virtual destructor you don't need to implement any of these, except for description these can be called from Lua

Reducing Coupling

• reducing coupling is a key software engineering principle
• improved reuse
• easier testing
  
• we want the OCR system to be pluggable
• e.g., combine different text/image, layout, text recognition components
• therefore: pass as little information as possible between modules
• e.g., text line recognizer should work if it is given a readable input, regardless of resolution etc.
  
• we want the OCR system to require little tuning
• therefore: reduce the number of parameters for each module to the bare minimum
• try to estimate parameters dynamically from input data

IDocumentCleanup

    struct ICleanupBinary : IComponent {
        virtual void cleanup(bytearray &out,bytearray &in) = 0;
    };

IDocumentCleanup Implementation

struct SampleCleanup : ICleanupBinary {
    const char *description() {
        return "SampleCleanup";
    }
    void cleanup(bytearray &out,bytearray &in) {
        float frac = 1.5;
        floatarray hist(100);
        intarray runpeaks;
        runlength_histogram(hist,in);
        dshow1d(hist); dwait();
        peaks(runpeaks,hist,2,10,3.0);
        int threshold = 2;
        if(runpeaks.length()>0) threshold = int(runpeaks[0] * frac);
        intarray blobs;
        blobs.copy(in);
        label_components(blobs);
        remove_small_components(blobs,threshold,threshold);
        greater(blobs,0,0,255);
        out.copy(blobs);
    }
};

IDocumentCleanup Implementation (Parameters)

struct SampleCleanup : ICleanupBinary {
    float frac;
    SampleCleanup() {
        frac = 1.5;
    }
    const char *description() {
        return "SampleCleanup";
    }
    void set(const char *name,double value) {
        if(!strcmp(name,"frac")) frac = value;
        else throw "unknown parameter setting";
    }
    double getd(const char *name) {
        if(!strcmp(name,"frac")) return frac;
        else throw "unknown parameter setting";
    }
    void cleanup(bytearray &out,bytearray &in) {
        ...
    }
};

IDocumentCleanup Implementation (Argument Checking)

struct SampleCleanup : ICleanupBinary {
    ...
    void cleanup(bytearray &out,bytearray &in) {
        optional_check_background_is_lighter(in);
        CHECK_ARG(contains_only(in,0,255));
        ...
    }
};