How to improve tess-two OCR performace on Android?

[Chinwei Chang]

I'm trying to make an android application that do real time ocr and then put the result text on the corresponding text areas.

The libraries I used is:

    OpenCV4android 3.1.0.

    tess-two(https://github.com/rmtheis/tess-two)

The target language is: English.(I use eng.traineddata)

I've been testing it on my tablet(asus memo pad7). 

The text-detection part is good enough, but the ocr part is no better than gibberish and the frame rate drops.

I want to know if there is a method to improve

I did searched several articles to find a way to improve ocr result, but none of these worked.

I'm very confident that the real time ocr is possible, because 2 years ago I saw this video

I've been studying his code for over 6 months and finally made my way here.

But there is a striking difference between the OCR result of my app and his work.

I will keep trying to contact the author.

But he is now starting a new company at Japan and it seems that he will have any time to deal with my problem.

Therefore I post this question here, hoping anyone could give me some help.

The following is my code to detect text and do ocr:

public class TextDetector implements Filter{

Bitmap bmp=null;

TessBaseAPI baseApi = new TessBaseAPI();

public String ocr(Mat tmp){

        baseApi.init("/mnt/sdcard/tesseract/", "eng");

        baseApi.setPageSegMode(TessBaseAPI.PageSegMode.PSM_SINGLE_LINE);

        bmp=Bitmap.createBitmap(tmp.cols(), tmp.rows(), Bitmap.Config.ARGB_8888);

        Utils.matToBitmap(tmp, bmp);

        baseApi.setImage(bmp);

        return baseApi.getUTF8Text();

}

@Override

public void apply(Mat src, Mat dst) {

if (dst != src) {

src.copyTo(dst);

}

//this text detection method is translated from http://stackoverflow.com/questions/23506105/extracting-text-opencv

   Mat img_gray,img_sobel, img_threshold, element;

   //turn source image to gray scale

   img_gray=new Mat();

   Imgproc.cvtColor(src, img_gray, Imgproc.COLOR_RGB2GRAY);

   //sobel edge detection

   img_sobel=new Mat();

   Imgproc.Sobel(img_gray, img_sobel, CvType.CV_8U, 1, 0, 3, 1, 0,Core.BORDER_DEFAULT);

   //thresholding

   img_threshold=new Mat();

   Imgproc.threshold(img_sobel, img_threshold, 0, 255, Imgproc.THRESH_OTSU+Imgproc.THRESH_BINARY);

   //removing noise

   element=new Mat();

   element = Imgproc.getStructuringElement(Imgproc.MORPH_RECT, new Size(17, 3) );

   Imgproc.morphologyEx(img_threshold, img_threshold, Imgproc.MORPH_CLOSE, element);

   //extract contours

   List<MatOfPoint>  contours=new ArrayList<MatOfPoint>();

   Mat hierarchy = new Mat();

   Imgproc.findContours(img_threshold, contours, hierarchy, 0, 1);

   List<MatOfPoint> contours_poly=new ArrayList<MatOfPoint>(contours.size());

   //for opencv4android only. you have to turn the List<MatOfPoint> into MatOfPoint2f

   //to run the approxPolyDP function. then turn it back into List<MatOfPoint>

   contours_poly.addAll(contours);

   MatOfPoint2f mMOP2f1,mMOP2f2;

   mMOP2f1=new MatOfPoint2f();

   mMOP2f2=new MatOfPoint2f();

   //

   for( int i = 0; i < contours.size(); i++ )

       if (contours.get(i).toList().size()>100)

       { 

        contours.get(i).convertTo(mMOP2f1, CvType.CV_32FC2);

        Imgproc.approxPolyDP(mMOP2f1,mMOP2f2, 3, true );

        mMOP2f2.convertTo(contours_poly.get(i), CvType.CV_32S);

           Rect appRect=Imgproc.boundingRect(contours_poly.get(i));

           if (appRect.width>appRect.height) 

           {

            Imgproc.rectangle(dst, new Point(appRect.x,appRect.y) ,new Point(appRect.x+appRect.width,appRect.y+appRect.height), new Scalar(255,0,0));

               Log.i(null, "cut Roi");

            Mat roiMat=new Mat(src,appRect);            

            Log.i(null, "OCR and puttext");

            Imgproc.putText(dst,ocr(roiMat), new Point(appRect.x,appRect.y),Core.FONT_HERSHEY_SIMPLEX, 1, new Scalar(0,255,0));

            Log.i(null, "Complete!");

           

           }

       }

   

}

}