Depending on the quality of your source images, I think it'd be reasonable to scale them down in order for letters to have the height of 40 pixels or so. In that way Tesseract will just have to do a bit less work - scan lesser pixels and construct shorter glyph outlines.
The accuracy may suffer even for such a considerable char height (90 is certainly more than enough) if you have significant discrepancies between training and source images. You should try to pass to Tesseract images having as similar as possible thickness and orientation. To achieve this, you need to pre-process images to get them look alike with respect to lighting conditions, contrast, blur amount, physical dimensions; rectify perspective distortion, etc. And of course, always use the same binarization procedure with the same parameter set, or at least giving predictably similar results for a range of your source images. Btw, using Otsu thresholding prior to passing images to Tesseract is useless as Otsu is a binarization procedure employed by Tesseract itself. Except if you do Otsu with your own special parameter set and then pass a 1-bit image. Next, you should train Tesseract having in mind that ideally there should be around 20 samples of each char. You shouldn't be striving to train using as many as possible char sizes - regardless of the size, Tesseract scales character "models" up or down to the same internal dimensions. But if your source char sizes differ - that's no problem, they'll do. Provide real images (probably pre-processed) images for training, not manually compiled ones. What can be done to further improve the speed and accuracy - process your images char by char, bypassing Tesseract's layout analysis. This approach also perfectly allows to use char-position-specific whitelists (letters, digits) for even more speedup and precision. Everything related to Tesseract's dictionary facility is totally irrelevant here. You'd better provide entirely empty files for your "traineddata". HTH Warm regards, Dmitri Silaev On Tue, Mar 29, 2011 at 8:17 AM, Andres <[email protected]> wrote: > ...required. > > Hello people, > > I'm develping a licence plate recognition system from long ago and I still > have to improve the use of Tesseract to make it usable. > > My first concern is about speed: > After extracting the licence plate image, I get an image like this: > > https://docs.google.com/leaf?id=0BxkuvS_LuBAzNmRkODhkYTUtNjcyYS00Nzg5LWE0ZDItNWM4YjRkYzhjYTFh&hl=en&authkey=CP-6tsgP > > As you may see, there are only 6 characters (tess is recognizing more > because there are some blemishes over there, but I get rid of them with some > postprocessing of the layout of the recognized chars) > > In an Intel I7 720 (good power, but using a single thread) the tesseract > part is taking something like 230 ms. This is too much time for what I need. > > The image is 500 x 117 pixels. I noted that when I reduce the size of this > image the detection time is reduced in proportion with the image area, which > makes good sense. But the accuracy of the OCR is poor when the characters > height is below 90 pixels. > > So, I assume that there is a problem with the way I trained tesseract. > > Because the characters in the plates are assorted (3 alphanumeric, 3 > numeric) I trained it with just a single image with all the letters in the > alphabet. I saw that you suggest large training but I imagine that that > doesn't apply here where the characters are not organized in words. Am I > correct with this ? > > So, for you to see, this is the image with what I trained Tesseract: > > https://docs.google.com/viewer?a=v&pid=explorer&chrome=true&srcid=0BxkuvS_LuBAzODc1YjIxNWUtNzIxMS00Yjg3LTljMDctNDkyZGIxZWM4YWVm&hl=en&authkey=CMXwo-AL > > In this image the characters are about 55 pixels height. > > Then, for frequent_word_list and words_list I included a single entry for > each character, I mean, something starting with this: > > A > B > C > D > ... > > Do you see something to be improved on what I did ? Should I perhaps use a > training image with more letters, with more combinations ? Will that help > somehow ? > > Should I include in the same image a copy the same character set but with > smaller size ? In that way, will I be able to pass Tesseract smaller images > and get more speed without sacrificing detection quality ? > > > On the other hand, I found some strange behavior of Tesseract about which I > would like to know a little more: > In my preprocessing I tried Otsu thresholding > (http://en.wikipedia.org/wiki/Otsu%27s_method) and I visually got too much > better results, but surprisingly for Tesseract it was worse. It decreased > the thickness of the draw of the chars, and the chars I used to train > Tesseract were bolder. So, Tesseract matches the "boldness" of the > characters ? Should I train Tesseract with different levels of boldness ? > > I'm using Tesseract 2.04 for this. Do you think that some of these issues > will go better by using Tess 3.0 ? > > > Thanks, > > Andres > > > > > > > -- > You received this message because you are subscribed to the Google Groups > "tesseract-ocr" group. > To post to this group, send email to [email protected]. > To unsubscribe from this group, send email to > [email protected]. > For more options, visit this group at > http://groups.google.com/group/tesseract-ocr?hl=en. > -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To post to this group, send email to [email protected]. To unsubscribe from this group, send email to [email protected]. For more options, visit this group at http://groups.google.com/group/tesseract-ocr?hl=en.
