改善tesseract结果（pytesseract）

Question

I have been downloading tiles from a TMS-like server displaying some "google tiles" with geodatas. Those datas are french townships, set in some particular colors regarding to the map's legend.

I have written an algorithm which mainly uses PIL to process tiles (as pictures) before presenting those to tesseract (using pytesseract). In the end, knowing the tile's position (and therfore knowing which townships may be in the area), I use fuzzywuzzy process.extractBests to try to identify which township has been found.

So far, some pictures which don't handle any result from tesseract seem relatively fine to me (though it is not perfect, you can distincly read the french name "Sainte Honorine de Ducy") :

I should precise that in this case, the original tile is around 1500x3000 pixels (I have already been extending the tile's size).

I have also modified pytesseract to pass the 'bazaar' keyword which was mentionned in the doc plus custom 'user-words' containing townships from the area. That beeing said, I could only find "bazaar" reference in tesseract 1 documentation , without anything better than a link in newest documentations. In fact, I seem to remember a post somewhere about it beeing a mistake in documentation... For what it's worth, it doesn't seem to change the results here.

Do you have any suggestions ? In particular, do you think the picture's quality would be considered good enough to expect solid results ?

I know almost nothing about training tesseract on my own with these particular font. Considering that (and that I don't manage the datasource, that I don't even know what font is used...), I hope that you may have better suggestions than taking this (huge) leap...

PS : I know I maybe shouldn't have posted this question whithout any code, but I'm more in a need for global guidance here... I will post any recquired code anyway !

Answer 1

I think that the problem is that the text is too small compared to the image size.

You should apply some more image transformations to find a more exact area where the text is located, try something as morphological transformations and then find the contours of the area with the text. Take a look also as this tutorial, it's with OpenCV.

I tried to crop the image with GIMP and then resized it to make a little bigger:

The result with pytesseract is:

Saiptnmnorine-de-Ducy

that is acceptable, with some other processings with fuzzywuzzy you could get the right name.