Hello all, I am training Tesseract to recognize specific images taken by a cell phone camera. I plan to create a new "language" and 2 new fonts for this training. In theory, this should be very simple and easy to do, but in fact I got lower accuracy with my new .traineddata than with the standard eng.traineddata. The more images I used for my training, the lower the accuracy I got.
The texts in the images varied in boldness and noise. I've tried correcting them with ImageMagick (300 density, black and white). <https://lh4.googleusercontent.com/-XacxrUljrKE/U8jApPMb5rI/AAAAAAAAAIY/zAGOKMT_T7s/s1600/ktp.general.exp81.jpg> <https://lh4.googleusercontent.com/-AxytQem9yW0/U8jAf5jwU6I/AAAAAAAAAII/AJA_AKsUVvI/s1600/ktp.general.exp01.jpg> <https://lh4.googleusercontent.com/-XCff6pZhEuk/U8jAihIh8-I/AAAAAAAAAIQ/5WQEYdnS0Ls/s1600/ktp.general.exp31.jpg> Notice that image in the middle (no.2) has bolder letters than the others. The white area is cleared out because of noise. Here's what I've done: 1. Adding a word-dawg file including the common words in the images. 2. Adding a unicharambigs file including the common mistakes like VV for W 3. Selecting the good letter model. The noisy letters were not included in the training. Please suggest what I should do more to get higher accuracy. Thanks in advance. Regards, Victoria -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/tesseract-ocr. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/723c2f3a-f2aa-411e-b99e-8ca1d65d6c66%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
