Question

Using the javaocr framework from sourceforge. Trying to scan letters from a image, and training the system to recognize them.

Getting this exception when loading trainer:

java.io.IOException: Expected to decode 26 characters but actually decoded 33 characters in training: /Developer/MAckan/bin/LETTERS/trainLetters.PNG
    at net.sourceforge.javaocr.ocrPlugins.mseOCR.TrainingImageLoader.load(TrainingImageLoader.java:111)

My code is like this:

loader.load(this,ClassLoader.getSystemResource("LETTERS/trainLetters.PNG").getPath(), new CharacterRange('A', 'Z'), images);

Another question is how to get it to train Scandinavian letters. If I enter a range A-Ö it expects 150 characters.

Then when I scan I try and scan a line in the image at the time:

scanner.addTrainingImages(images);

        final CharacterRange[] cr = new CharacterRange[1];
        cr[0] = new CharacterRange('A', 'Z');

        // get the first line of letters
        final int x1 = 0;
        final int y1 = 130;
        final int x2 = 640;
        final int y2 = 170;


        for (int i = 0; i < 15; i++) {
        final String text = scanner.scan(boardImage, x1, y1 + (i * 40), x2,
                    y2 + (i * 40), cr);
            System.out.println("scanned " + text);
        }

And I actually get output, but not the output I expect…
Anyone have experience with the javaocr framework?

Update:
Solved the training issue. The training image was missing a couple of charachters and Scandinavian is not supported (?). Still getting strange output.

Update2:
Solved the entire issue with writing my own comparison instead. I did some manipulation of the images (reduced colors and transperency) and compared pixel by pixel and returned a diff against alafabet images. The lowest diff “wins”. Works for this particular case, but I am still interested in getting OCR running.

Thanks.

/A

Answer 1

Solved the entire issue with writing my own comparison instead. I did some manipulation of the images (reduced colors and transperency) and compared pixel by pixel and returned a diff against alafabet images. The lowest diff “wins”. Works for this particular case, but I am still interested in getting OCR running.

Thanks everyone for contributing.

/A