improve Tesseract performance with OpenCV on Android

Question

I am working on a Android application using real-time OCR. I using OpenCV and Tesseract Library. But the performance is very poor, even on my Galaxy SIII. There are any methods to improve the performance? It is my code:

    Mat mGray = new Mat();
capture.retrieve(mGray);
Bitmap bmp = Bitmap.createBitmap(mGray.cols(), mGray.rows(), Bitmap.Config.ARGB_8888);
tessBaseApi.setImage(bmp);
String recognizedText = tessBaseApi.getUTF8Text();
Log.i("Reg", recognizedText);

Will the speed of tesseract OCR be reduced by passing bitmap to the Tesseract API? What pre-processing should I perform before passing to the Tesseract API?

Hey @QuiLlHoN Did you find any solutions for such a slow performance? I am encountering the same problem :/ — Vucko, Oct 17 '16 at 14:58

score 2 · Answer 1 · answered Apr 21 '13 at 08:57

2

One thing to try is to binarize the image using adaptive thresholding (adaptiveThreshold in OpenCV).

answered Apr 21 '13 at 08:57

ojs

285
1
13

1

Tesseract already performs this internally. It uses Otsu's Thresholding. https://code.google.com/p/tesseract-ocr/source/browse/ccstruct/otsuthr.cpp – Raghav Feb 12 '15 at 22:51
Otsu's method uses a single global threshold. This does not work as well as adaptive thresholding if the lighting is not completely uniform (see http://docs.opencv.org/trunk/doc/py_tutorials/py_imgproc/py_thresholding/py_thresholding.html#adaptive-thresholding for examples) – ojs Feb 17 '15 at 12:39
Indeed Otsu does not work as great, but it is still very fast. So while this might solve precision issues, it does not affect performance greatly. – Dmitry Zaytsev Apr 10 '15 at 09:24

score 1 · Answer 2 · answered Oct 04 '12 at 17:45

1

You can have Tesseract only do the recognition pass 1, so that it skips passes 2 through 9, when it calls recog_all_words().

Change the following line in baseapi.cpp and rebuild your Tesseract library project:

if (tesseract_->recog_all_words(page_res_, monitor, NULL, NULL, 0)) {

Change it to:

if (tesseract_->recog_all_words(page_res_, monitor, NULL, NULL, 1)) {

answered Oct 04 '12 at 17:45

rmtheis

5,992
12
61
78

I edited the code and rebuilt the library. But the speed still very slow. – QuiLl HoN Oct 06 '12 at 08:23

score 0 · Answer 3 · answered Oct 04 '12 at 09:00

0

Some things that might make it faster are:

Select a smaller region from mGray where your text is, before createBitmap - so the more heavy methods that follow process a smaller image.
Changing Bitmap.Config.ARGB_8888 to Bitmap.Config.RGB_565 - your image is grayscale, it will not need a ARGB bitmap.

answered Oct 04 '12 at 09:00

Rui Marques

8,567
3
60
91

The TessBaseAPI only accept ARGB_8888 image. There are any algorithms for finding the text region? Thank you. – QuiLl HoN Oct 06 '12 at 03:45

score 0 · Answer 4 · answered Jan 19 '17 at 11:20

Use multithreading, but be aware to create one instance per thread for TessBaseAPI. Don't share them between different threads. Create N threads (N >= number of cores), and java will make sure that you speed up at least the number of cores times.

What I do is creating N threads which create TessBaseAPI objects in their own context (in the run method) and wait for OCR requests in a loop until interrupted.

    ...
    ...
    @Override
    public void run() {

      TessBaseAPI tessBaseApi = new TessBaseAPI();

      tessBaseApi.init(Ocrrrer.DATA_PATH, "eng");

      setTessVariable(tessBaseApi, "load_system_dawg", "0");
      setTessVariable(tessBaseApi, "load_freq_dawg", "0");
      setTessVariable(tessBaseApi, "load_unambig_dawg", "0");
      setTessVariable(tessBaseApi, "load_punc_dawg", "0");
      setTessVariable(tessBaseApi, "load_number_dawg", "0");
      setTessVariable(tessBaseApi, "load_fixed_length_dawgs", "0");
      setTessVariable(tessBaseApi, "load_bigram_dawg", "0");
      setTessVariable(tessBaseApi, "wordrec_enable_assoc", "0");
      setTessVariable(tessBaseApi, "tessedit_enable_bigram_correction", "0");
      setTessVariable(tessBaseApi, "assume_fixed_pitch_char_segment", "1");
      setTessVariable(tessBaseApi, TessBaseAPI.VAR_CHAR_WHITELIST, "1234567890ABCDEFGHIJKLMNOPQRSTUVWXYZ<");

      Log.d(TAG, "Training file loaded");


      while (!interrupted()) {
        reentrantLock.lock();
        try {
          Log.d(TAG, this.getName() + " wait for OCR");
          jobToDo.await();
          Log.d(TAG, this.getName() + " input arrived. Do OCR");
          this.ocrResult = doOcr(tessBaseApi);
          ocrDone.signalAll();
        } catch (InterruptedException e) {
          return;
        } finally {
          try {
            reentrantLock.unlock();
          } catch (Exception ex) {
          }
        }
      }

    }
    ...
    ...

You can see that the tessBaseApi object is local to the run method, hence absolutely not shared.

improve Tesseract performance with OpenCV on Android

4 Answers4

Linked