I am working on a project that involves OCRing an extremely large number of PDFs. Each PDF is an image of a normal, office-type document. Each PDF page is stored a separate file. Is there any way to do this faster (i.e. with lower compute time) than my current approach? With a project this size, any small performance improvement would yield a large reduction in computing cost/time. Currently, the call to Tesseract takes about 2-3 orders of magnitude more time than all other processing steps in the project. For each file (page.pdf), I first convert it to Tiff (image.tiff), like so. I stripped the Python off of this call for clarity. ghostscript -o image.tiff -q -r300x300 -sCompression=none -sDEVICE=tiffgray page.pdf Then I run essentially the following procedure. def ocr( image ): import tesseract, cv2.cv as cv # Load Tesseract api = tesseract.TessBaseAPI() api.Init(".","eng",tesseract.OEM_DEFAULT) api.SetPageSegMode(tesseract.PSM_AUTO) # Load and OCR the image file cv_image = cv.LoadImage(image, cv.CV_LOAD_IMAGE_GRAYSCALE) tesseract.SetCvImage(cv_image, api) # Return the text return api.GetUTF8Text() I've simplified a bit for clarity. I don't actually separately load Tesseract for each file, for instance. I handle exceptions, etc. However, the differences don't really matter, since 99.9% of the time used is in the api.GetUTF8Text() call. Any ideas to speed it up? Continue reading...
