But we were trying to find a crop in the first For some images, Tesseract misses the text completely.Ĭropping fixes the problem. These should be the areas that we crop to! But this is a bit of a chickenĪnd the egg problem. I ran the image through Tesseract to find areas which contained letters. I tried several other approaches which didn’t work as well.
OPENCV TEXT ON IMAGE INSTALL
You’ll need to install OpenCV, numpy and PIL If you want to try using this procedure to crop your own images, you can find
Depending on how youĬount, I’d estimate that it gets a perfect crop on about 98% of the images, and This procedure worked well for my particular application. On what’s important, rather than the noise. The number of pixels, with no loss of text! This will help any OCR tool focus We do this repeatedly until there are only a few connected To do this, we apply binaryĭilation to the de-bordered edge image. Problem by finding individual chunks of text. The saving grace is that most crops don’t make much sense. Where W and H are the width and height of the image. The set of all possible crops is quite large: W 2 H 2, Score, the harmonic mean of precision and recall.
To smudges or marks on the original page.Īt this point, we’re looking for a crop (x1, y1, x2, y2) which: What we’re left with is an image with the text and possibly some other bits due With polygons for the borders, it’s easy to black out everything outside them. Noticed that it performed much better on the Milstein images when I manuallyĬropped them down to just the text regions first: Source project developed over the past 20+ years at HP and Google. The most famous OCR program is Tesseract, a remarkably long-lived open Page layout analysis, a much less glamorous problem, is at least as important But it’s a dirty secret of the trade that When you hear “OCR”, you might think about fancy Machine Learning OCR programs typically have to do some sort of page-layout analysis toįind out where the text is and carve it up into individual lines andĬharacters. There are ~34,000 images: too many to affordably turk.The images are ~4x the resolution shown here (2048px tall).The image is slightly rotated from vertical.But the typewriter font isn’t always consistent across the collection. The text is written with a tyepwriter, so it’s monospace.Only a small portion of the image contains text.There’s a black border around the whole image, gray backing paper and then white paper with text on it.Public Member Functions inherited from cv::text::BaseOCRĬreate (const char *datapath=NULL, const char *language=NULL, const char *char_whitelist=NULL, int oem= OEM_DEFAULT, int psmode= PSM_AUTO)Ĭreates an instance of the OCRTesseract class. SetWhiteList (const String &char_whitelist)=0 Run ( InputArray image, InputArray mask, int min_confidence, int component_level=0) Run ( InputArray image, int min_confidence, int component_level=0) Run ( Mat &image, Mat &mask, std::string &output_text, std::vector *component_rects=NULL, std::vector *component_texts=NULL, std::vector *component_confidences=NULL, int component_level=0) CV_OVERRIDE Recognize text using the tesseract-ocr API. Run ( Mat &image, std::string &output_text, std::vector *component_rects=NULL, std::vector *component_texts=NULL, std::vector *component_confidences=NULL, int component_level=0) CV_OVERRIDE
0 kommentar(er)
