hocr | facilitate post-OCR data processing | Computer Vision library

If the file name contains spaces, then you need to enclose the name in quotation marks.

The code that handles the X-Tika-OCR and X-Tika-PDF headers is TikaResource.processHeaderConfig.

Those header suffixes and values are then mapped onto the TesseractOCRConfig and PDFParserConfig configuration objects via reflection.

So, to see what X-Tika headers you can set, look up the options on the config class you want to tweak things on (Tesseract or PDF), then build the name, then set the header. If you are not sure what the option does, or what values it takes, look at the JavaDocs for the underlying setter method that will get called.

For eg setExtractInlineImages on PDF, that maps to X-Tika-PDFextractInlineImages

By inspecting the integration test code of TikaResourceTest, I realized an HTTP header was missing. The correct command should include the X-Tika-PDFOcrStrategy: ocr_only HTTP header. See more in the ocr & pdf parser docs

The command would thus be:

I would use cv2 and/or numpy.array

to convert light gray colors to white

You have the coordinates of the bounding box in every line.

From: Training Tesseract – Make Box Files

character, left, bottom, right, top, page

So for each character you get the character, followed by its bounding box characters, followed by the 0-based page number.

Replace pytesseract.run_tesseract() with pytesseract.pytesseract.run_tesseract().

Credit Nithin in the comments. Adding this as an answer to close it out.

Here's a simple approach

• Convert image to grayscale
• Otsu's threshold
• Dilate to connect contours
• Find contours and extract ROI for each word
• Perform OCR and remove word

After converting to grayscale, we Otsu's threshold to obtain a binary image

Next we invert the image and dilate to form a single contour for each word

From here we find contours and extract the ROI for each word. Here's the detected ROIs

We throw each ROI into Pytesseract OCR. If the OCR result is a word we want to remove, we simply "delete" the word by filling in the ROI with white and replace it in the original image

With

It turns out that the reason that pytesseract was not able to find the temporary output files were that they were being stored with extensions other than .txt or .box (they were .hocr files). From the source code, these are the only types of tesseract output files supported by pytesseract (or more like 'looked for' by pytesseract). The relevant snippets from the source are below:

input_file_name = '%s.bmp' % tempnam() output_file_name_base = tempnam() if not boxes: output_file_name = '%s.txt' % output_file_name_base else: 123 output_file_name = '%s.box' % output_file_name_base

if status: errors = get_errors(error_string) raise TesseractError(status, errors) 135 f = open(output_file_name, 'rb')

Looking at the pytesseract's github pulls, it seems like support for other output types is planned but not yet implemented (the source code I used to show why .hocr file were appearing to not be found was copy/pasted from the pytesseract master branch).

Until then, I made some hackish changes to the pytesseract script to support multiple file types.

This version does not set an extension for the output file (since tesseract does that automatically) and looks through the directory that pytesseract stores its temp output files to and looks for the file that starts with the output file name (up to the first '.' character) assigned by pytesseract (without caring about the extension):

Figured out a (gross) way to do it using xpath.