eihli
/
image-table-ocr
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
main
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from '3b31888a55'
${ noResults }
image-table-ocr/ocr_tables

14 lines
575 B
Plaintext
Raw Normal View History Unescape Escape

Fix bugs and improve accuracy Files in the ocr_to_csv module need to be named in a certain way. Specify that and fix a bug, we need to have them sorted lexicographically. Don't dilate the characters in a cell in order to make a contiguous set of pixels that we can find a contour around. The problem with that is that you sometimes dilate too far and hit an image boundary and can't erode back in. If a cell wall border was remaining between the text and the image boundary, well now you're keeping that border line in the image. (Unless you remove it some other way. So that might be a valid option in the future.) The method we're using now instead is to group all contours together and create a bounding box around all of them. The problem with that is if there is any noise at all outside the text, we're grabbing it. Before, we were dilating and taking the largest contour, so we weren't including that noise. And we can't get rid of the noise with opening morph because it's sometimes pretty big noise and opening any bigger distorts the text so much that we lose accuracy in finding those boundaries. Also adds a shell script to simplify the plumbing of all these modules.
5 years ago
#!/bin/sh
PDF=$1
Clarify and expand on content around code
5 years ago
python -m table_ocr.pdf_to_images $PDF | grep .png > /tmp/pdf-images.txt
Fix bugs and improve accuracy Files in the ocr_to_csv module need to be named in a certain way. Specify that and fix a bug, we need to have them sorted lexicographically. Don't dilate the characters in a cell in order to make a contiguous set of pixels that we can find a contour around. The problem with that is that you sometimes dilate too far and hit an image boundary and can't erode back in. If a cell wall border was remaining between the text and the image boundary, well now you're keeping that border line in the image. (Unless you remove it some other way. So that might be a valid option in the future.) The method we're using now instead is to group all contours together and create a bounding box around all of them. The problem with that is if there is any noise at all outside the text, we're grabbing it. Before, we were dilating and taking the largest contour, so we weren't including that noise. And we can't get rid of the noise with opening morph because it's sometimes pretty big noise and opening any bigger distorts the text so much that we lose accuracy in finding those boundaries. Also adds a shell script to simplify the plumbing of all these modules.
5 years ago
cat /tmp/pdf-images.txt | xargs -I{} python -m table_ocr.extract_tables {} | grep table > /tmp/extracted-tables.txt
Clarify and expand on content around code
5 years ago
cat /tmp/extracted-tables.txt | xargs -I{} python -m table_ocr.extract_cells {} | grep cells > /tmp/extracted-cells.txt
cat /tmp/extracted-cells.txt | xargs -I{} python -m table_ocr.ocr_image {} --psm 7 -l table-ocr
Fix bugs and improve accuracy Files in the ocr_to_csv module need to be named in a certain way. Specify that and fix a bug, we need to have them sorted lexicographically. Don't dilate the characters in a cell in order to make a contiguous set of pixels that we can find a contour around. The problem with that is that you sometimes dilate too far and hit an image boundary and can't erode back in. If a cell wall border was remaining between the text and the image boundary, well now you're keeping that border line in the image. (Unless you remove it some other way. So that might be a valid option in the future.) The method we're using now instead is to group all contours together and create a bounding box around all of them. The problem with that is if there is any noise at all outside the text, we're grabbing it. Before, we were dilating and taking the largest contour, so we weren't including that noise. And we can't get rid of the noise with opening morph because it's sometimes pretty big noise and opening any bigger distorts the text so much that we lose accuracy in finding those boundaries. Also adds a shell script to simplify the plumbing of all these modules.
5 years ago
for image in $(cat /tmp/extracted-tables.txt); do
dir=$(dirname $image)
python -m table_ocr.ocr_to_csv $(find $dir/cells -name "*.txt")
done