image-table-ocr/pdf_table_extraction_and_ocr.org

# -*- org-image-actual-width: 500; -*-

#+TITLE: PDF Parsing
#+PROPERTY: header-args :session *Python*
#+STARTUP: inlineimages
#+OPTIONS: ^:nil H:4

#+BEGIN_COMMENT
Some notes about the header for those not familiar with Org Mode:

The property `header-args` with ~:session \*Python\*~ will cause all evaluated
source code blocks to be evaluated in the buffer named "\*Python\*", which is the
default buffer name for the buffer connected to an inferior python process. This
is useful for interactive development. It gives you a REPL to work with rather
than having to constantly evaluate source code blocks and view the results
output to try any change.

Another note along those lines is that when source code blocks are evaluated,
some unnecessary output is printed in the ~*Python*~ buffer. Adding ~:results
output~ to a code block will minimize that noise.
#+END_COMMENT

* Overview

This Python package provides utilities for extracting tabular data from PDF
files and images of tables.

Given an image that contains a table...

#+ATTR_HTML: :width 25%
[[file:resources/examples/example-page.png]]

Extract the the text into a CSV format...

#+BEGIN_EXAMPLE
PRIZE,ODDS 1 IN:,# OF WINNERS*
$3,9.09,"282,447"
$5,16.66,"154,097"
$7,40.01,"64,169"
$10,26.67,"96,283"
$20,100.00,"25,677"
$30,290.83,"8,829"
$50,239.66,"10,714"
$100,919.66,"2,792"
$500,"6,652.07",386
"$40,000","855,899.99",3
1,i223,
Toa,,
,,
,,"* Based upon 2,567,700"
#+END_EXAMPLE

The package is split into modules with narrow focuses.

- ~pdf_to_images~ uses Poppler and ImageMagick to extract images from a PDF.
- ~extract_tables~ finds and extracts table-looking things from an image.
- ~extract_cells~ extracts and orders cells from a table.
- ~ocr_image~ uses Tesseract to OCR the text from an image of a cell.
- ~ocr_to_csv~ converts into a CSV the directory structure that ~ocr_image~ outputs.

** Requirements

Tested with the following versions of the following packages

*** Python packages
- numpy 1.19.2
- opencv-python 4.4.0.44
- pytesseract 0.3.6

*** External
- ~pdfimages~ from Poppler 20.09.0
- ~tesseract~ 5.0.0
- ~mogfrify~ ImageMagick 7.0.10

** Contributing

This package was created in a [[https://en.wikipedia.org/wiki/Literate_programming][literate programming]] style with the help of [[https://orgmode.org/worg/org-contrib/babel/intro.html][Babel]].

The unfortunate downside is the obscurity of the tooling. It creates a bit of a
barrier for contributors who aren't already familiar with Emacs and Babel.

** Example usage

Here is an example of a shell script that uses each module to turn a pdf with a
table into CSV output.

Depending on your needs, you may not need all of these steps. If you already
have an image of a table, you can jum straight to extracting the cells.

Each piece is its own python module, so you can also simply import the pieces
you need into your own python projects and use them as needed.

#+NAME: ocr_tables
#+BEGIN_SRC shell :results none :tangle ocr_tables :tangle-mode (identity #o755)
#!/bin/sh

PDF=$1

python -m table_ocr.pdf_to_images $PDF | grep .png > /tmp/pdf-images.txt
cat /tmp/pdf-images.txt | xargs -I{} python -m table_ocr.extract_tables {}  | grep table > /tmp/extracted-tables.txt
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 {}
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
#+END_SRC

Any extra args you pass after the image path to ~python -m table_ocr.ocr_image~ will be passed directly to tesseract as options. If you don't pass anything, reasonable english defaults are used.

** Possible improvements

Detect text with the stroke-width-transform alogoritm. https://zablo.net/blog/post/stroke-width-transform-swt-python/index.html

* Preparing data

Not all pdfs need to be sent through OCR to extract the text content. If you can
click and drag to highlight text in the pdf, then the tools in this library
probably aren't necessary.

** Converting PDFs to images

This code calls out to [[https://manpages.debian.org/testing/poppler-utils/pdfimages.1.en.html][pdfimages]] from [[https://poppler.freedesktop.org/][Poppler]].

#+NAME: pdf-to-images
#+BEGIN_SRC python :results none
# Wrapper around the Poppler command line utility "pdfimages" and helpers for
# finding the output files of that command.
def pdf_to_images(pdf_filepath):
    """
    Turn a pdf into images
    Returns the filenames of the created images sorted lexicographically.
    """
    directory, filename = os.path.split(pdf_filepath)
    image_filenames = pdfimages(pdf_filepath)

    # Since pdfimages creates a number of files named each for there page number
    # and doesn't return us the list that it created
    return sorted([os.path.join(directory, f) for f in image_filenames])


def pdfimages(pdf_filepath):
    """
    Uses the `pdfimages` utility from Poppler
    (https://poppler.freedesktop.org/). Creates images out of each page. Images
    are prefixed by their name sans extension and suffixed by their page number.

    This should work up to pdfs with 999 pages since find matching files in dir
    uses 3 digits in its regex.
    """
    directory, filename = os.path.split(pdf_filepath)
    if not os.path.isabs(directory):
        directory = os.path.abspath(directory)
    filename_sans_ext = filename.split(".pdf")[0]

    # pdfimages outputs results to the current working directory
    with working_dir(directory):
        subprocess.run(["pdfimages", "-png", filename, filename.split(".pdf")[0]])

    image_filenames = find_matching_files_in_dir(filename_sans_ext, directory)
    logger.debug(
        "Converted {} into files:\n{}".format(pdf_filepath, "\n".join(image_filenames))
    )
    return image_filenames


def find_matching_files_in_dir(file_prefix, directory):
    files = [
        filename
        for filename in os.listdir(directory)
        if re.match(r"{}-\d{{3}}.*\.png".format(re.escape(file_prefix)), filename)
    ]
    return files
#+END_SRC

** Detecting image orientation and applying rotation.

Tesseract can detect orientation and we can then use [[https://www.imagemagick.org/script/mogrify.php][ImageMagick's mogrify]] to
rotate the image.

Here's an example of the output we get from orientation detection with
Tesseract.

#+BEGIN_EXAMPLE
➜  example/ tesseract --psm 0 example-000.png -
Page number: 0
Orientation in degrees: 90
Rotate: 270
Orientation confidence: 26.86
Script: Latin
Script confidence: 2.44
#+END_EXAMPLE

The following are some helpers to detect orientation of the images that Poppler
extracted and, if the images are rotated or skewed, use ImageMagick's `mogrify`
to correct the rotation. This makes OCR more straightforward.

#+NAME: fix-orientation
#+BEGIN_SRC python :results none
def preprocess_img(filepath, tess_params=None):
    """Processing that involves running shell executables,
    like mogrify to rotate.

    Uses tesseract to detect rotation.

    Orientation and script detection is only available for legacy tesseract
    (--oem 0). Some versions of tesseract will segfault if you let it run OSD
    with the default oem (3).
    """
    if tess_params is None:
        tess_params = ["--psm", "0", "--oem", "0"]
    rotate = get_rotate(filepath, tess_params)
    logger.debug("Rotating {} by {}.".format(filepath, rotate))
    mogrify(filepath, rotate)


def get_rotate(image_filepath, tess_params):
    """
    """
    tess_command = ["tesseract"] + tess_params + [image_filepath, "-"]
    output = (
        subprocess.check_output(tess_command)
        .decode("utf-8")
        .split("\n")
    )
    output = next(l for l in output if "Rotate: " in l)
    output = output.split(": ")[1]
    return output


def mogrify(image_filepath, rotate):
    subprocess.run(["mogrify", "-rotate", rotate, image_filepath])
#+END_SRC

* Detecting tables

This answer from opencv.org was heavily referenced while writing the code around
table detection:
https://answers.opencv.org/question/63847/how-to-extract-tables-from-an-image/.

It's much easier to OCR a table when the table is the only thing in the image.
This code detects tables in an image and returns a list of images of just the
tables, no surrounding text or noise.

The blurring, thresholding, and line detection is used here as well as later on
for cell extraction. They are good techniques for cleaning an image up in a way
that makes things like shape detection more accurate.

#+NAME: detect-tables
#+BEGIN_SRC python :results none :noweb yes
def find_tables(image):
    <<blur>>
    <<threshold>>
    <<lines-of-table>>
    contours, heirarchy = cv2.findContours(
        mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE,
    )

    MIN_TABLE_AREA = 1e5
    contours = [c for c in contours if cv2.contourArea(c) > MIN_TABLE_AREA]
    perimeter_lengths = [cv2.arcLength(c, True) for c in contours]
    epsilons = [0.1 * p for p in perimeter_lengths]
    approx_polys = [cv2.approxPolyDP(c, e, True) for c, e in zip(contours, epsilons)]
    bounding_rects = [cv2.boundingRect(a) for a in approx_polys]

    # The link where a lot of this code was borrowed from recommends an
    # additional step to check the number of "joints" inside this bounding rectangle.
    # A table should have a lot of intersections. We might have a rectangular image
    # here though which would only have 4 intersections, 1 at each corner.
    # Leaving that step as a future TODO if it is ever necessary.
    images = [image[y:y+h, x:x+w] for x, y, w, h in bounding_rects]
    return images
#+END_SRC

Here is an the an example of the result of the ~find_tables~ function.

#+HEADER: :post html-image-size(text=*this*, width="500px")
#+BEGIN_SRC python :noweb-ref test-detect-table :noweb strip-export :results none
import cv2
<<detect-tables>>
image_filename = "resources/examples/example-page.png"
image = cv2.imread(image_filename, cv2.IMREAD_GRAYSCALE)
image = find_tables(image)[0]
cv2.imwrite("resources/examples/example-table.png", image)
#+END_SRC

#+BEGIN_CENTER
#+ATTR_HTML: :width 250px
[[file:resources/examples/example-page.png]]

↓

#+ATTR_HTML: :width 250px
[[file:resources/examples/example-table.png]]
#+END_CENTER

** Improving accuracy

It's likely that some images will contain tables that aren't accurately
recognized by the code above. The code will then need to be made more robust.
But how will we know that changes to the code don't break the detection of
tables that were previously detected?

It might be good to add some type of test suite in the future that contains a
spec that matches a pdf with the pages and pixel coordinates of the detected
tables. The coordinates would need to have a range. Something like
"example-1.pdf, page-2.png, [450:470, 200:210, 800:820, 1270:1290]" where the
elements of the list are valid x, y, w, h ranges. So the test will pass if if
the x, y, width and height are anywhere in that range.

* OCR tables

Tesseract does not perform well when run on images of tables. It performs best
when given a single line of text with no extra noise.

Therefore, our next task is to find and extract the bounding box of each cell in
the table. Run tesseract on each cell. Print a comma seperated output.

We'll start with an image shown at the end of the previous section.

** Training Tesseract

Tesseract is used for recognizing characters. It is not involved in extracting the tables from an image or in extracting cells from the table.

It's a very good idea to train tesseract. Accuracy will improve tremendously.

Clone the tesstrain repo at [[https://github.com/tesseract-ocr/tesstrain]].

Run the [[ocr_tables][~ocr_tables~]] script on a few pdfs to generate some training data. That
script outputs pairs of ~.png~ and ~.gt.txt~ files that can be used by
tesstrain.

Make sure the ~.gt.txt~ files contain an accurate recognition of the
corresponding image. Since the first few runs will be untrained, you'll probably
need to fix up a few of the text files.

Once they are accurate, move them to a new subdirectory of the tesstrain repo;
~tesstrain/data/<model-name>-ground-truth/~.

You'll also need to clone the ~tessdata_best~ repo,
[[https://github.com/tesseract-ocr/tessdata_best]] and the
https://github.com/tesseract-ocr/langdata to use as the start of the
training model.

I'm actually not sure how much the punctuation and numbers from ~langdata~ help.
I didn't keep accurate records while playing with the training, I don't
thoroughly understand it, and it's not profitable for me to explore it at the
moment. It worked for my purposes and that has been good enough.

#+BEGIN_EXAMPLE
make training MODEL_NAME=table-ocr START_MODEL=eng TESSDATA=~/src/tessdata_best PUNC_FILE=~/src/langdata/eng/eng.punc NUMBERS_FILE=~/src/langdata/eng/eng.numbers
#+END_EXAMPLE

Once the training is complete, there will be a new file
~tesstrain/data/<model-name>.traineddata~. Copy that file to the directory
Tesseract searches for models. On my machine, it was ~/usr/local/share/tessdata/~.

*** Training tips

Here is a tip for quickly creating training data.

The output of the ~ocr_cells~ script will be a directory named ~ocr_data~ that
will have two files for each cell. One file is the image of the cell and the
other file is the OCR text.

You'll want to compare each image to its OCR text to check for accuracy. If
the text doesn't match, you'll want to update the text and add the image to the
training data.

The fastest way to do this is with ~feh~.
~feh~ lets you view an image and a caption at the same time and lets you edit
the caption from within ~feh~.
~feh~ expects the captions to be named ~<image-name>.txt~, so use a little
shell-fu to do a quick rename.

#+BEGIN_SRC shell :eval no
for f in *.txt; do f1=$(cut -d"." -f1 <(echo $f)); mv $f ${f1}.png.txt; done
#+END_SRC

Then run ~feh -K .~ to specify the current directory as the caption directory.
This will open a window with the first image in the directory and its caption.

Press ~c~ to edit the caption (if needed) and ~n~/~p~ to move to the
next/previons images. Press ~q~ to quit.

When finished, rename the files back to the filename structure that Tesseract
looks for in training.

#+BEGIN_SRC shell :eval no
for f in *.txt; do f1=$(cut -d"." -f1 <(echo $f)); mv $f ${f1}.gt.txt; done
#+END_SRC

** Blur

Blurring helps to make noise less noisy so that the overall structure of an
image is more detectable.

That gray row at the bottom is kind of noisy. If we don't somehow clean it up,
OpenCV will detect all sorts of odd shapes in there and it will throw off our
cell detection.

Cleanup can be accomplished with a blur followed by some thresholding.

#+BEGIN_SRC python :noweb-ref blur :results none
BLUR_KERNEL_SIZE = (17, 17)
STD_DEV_X_DIRECTION = 0
STD_DEV_Y_DIRECTION = 0
blurred = cv2.GaussianBlur(image, BLUR_KERNEL_SIZE, STD_DEV_X_DIRECTION, STD_DEV_Y_DIRECTION)
#+END_SRC

#+HEADER: :post html-image-size(text=*this*, width="500px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
image = ~cv2.imread("resources/examples/example-table.png", cv2.IMREAD_GRAYSCALE)
<<blur>>
cv2.imwrite("resources/examples/example-table-blurred.png", blurred)
#+END_SRC

#+ATTR_HTML: :width 500px :height 100%
[[file:resources/examples/example-table-blurred.png]]

** Threshold

We've got a bunch of pixels that are gray. Thresholding will turn them all
either black or white. Having all black or white pixels lets us do morphological
transformations like erosion and dilation.

#+BEGIN_SRC python :noweb-ref threshold :results none
MAX_COLOR_VAL = 255
BLOCK_SIZE = 15
SUBTRACT_FROM_MEAN = -2

img_bin = cv2.adaptiveThreshold(
    ~blurred,
    MAX_COLOR_VAL,
    cv2.ADAPTIVE_THRESH_MEAN_C,
    cv2.THRESH_BINARY,
    BLOCK_SIZE,
    SUBTRACT_FROM_MEAN,
)
#+END_SRC

#+HEADER: :post html-image-size(text=*this*, width="500px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
<<threshold>>
cv2.imwrite("resources/examples/example-table-thresholded.png", img_bin)
#+END_SRC

#+ATTR_HTML: :width 500px :height 100%
[[file:resources/examples/example-table-thresholded.png]]

** Finding the vertical and horizontal lines of the table

#+BEGIN_SRC python :noweb-ref lines-of-table :results none
vertical = horizontal = img_bin.copy()
SCALE = 5
image_width, image_height = horizontal.shape
horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (int(image_width / SCALE), 1))
horizontally_opened = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN, horizontal_kernel)
vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, int(image_height / SCALE)))
vertically_opened = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN, vertical_kernel)

horizontally_dilated = cv2.dilate(horizontally_opened, cv2.getStructuringElement(cv2.MORPH_RECT, (40, 1)))
vertically_dilated = cv2.dilate(vertically_opened, cv2.getStructuringElement(cv2.MORPH_RECT, (1, 60)))

mask = horizontally_dilated + vertically_dilated
#+END_SRC

Note: There's a wierd issue with the results of the example below when it's
evaluated as part of an export or a full-buffer evaluation. If you evaluate the
example by itself, it looks the way it's intended. If you evaluate it as part of
an entire buffer evaluation, like during export, it's distorted.

#+HEADER: :post html-image-size(text=*this*, width="500px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
<<lines-of-table>>
cv2.imwrite("resources/examples/example-table-lines.png", mask)
#+END_SRC

#+ATTR_HTML: :width 500px
[[file:resources/examples/example-table-lines.png]]

** Finding the contours

Blurring and thresholding allow us to find the lines. Opening the lines allows
us to find the contours.

An "Opening" is an erosion followed by a dilation. Great examples and
descriptions of each morphological operation can be found at
[[https://docs.opencv.org/trunk/d9/d61/tutorial_py_morphological_ops.html][https://docs.opencv.org/trunk/d9/d61/tutorial_py_morphological_ops.html]].

#+BEGIN_QUOTE
Contours can be explained simply as a curve joining all the continuous points
(along the boundary), having same color or intensity. The contours are a useful
tool for shape analysis and object detection and recognition.
#+END_QUOTE

We can search those contours to find rectangles of certain size.

To do that, we can use OpenCV's ~approxPolyEP~ function. It takes as arguments
the contour (list of contiguous points), and a number representing how different
the polygon perimeter length can be from the true perimeter length of the
contour. ~0.1~ (10%) seems to be a good value. The difference in perimeter
length between a 4-sided polygon and a 3-sided polygon is greater than 10% and
the difference between a 5+ sided polygon and a 4-sided polygon is less than
10%. So a 4-sided polygon is the polygon with the fewest sides that leaves the
difference in perimeter length within our 10% threshold.

Then we just get the bounding rectangle of that polygon and we have our cells!

We might need to do a little more filtering of those rectangles though. We might
have accidentally found some noise such as another image on the page or a title
header bar or something. If we know our cells are all within a certain size (by
area of pixels) then we can filter out the junk cells by removing cells
above/below certain sizes.

#+NAME: bounding-rects
#+BEGIN_SRC python :results none
contours, heirarchy = cv2.findContours(
    mask, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE,
)

perimeter_lengths = [cv2.arcLength(c, True) for c in contours]
epsilons = [0.05 * p for p in perimeter_lengths]
approx_polys = [cv2.approxPolyDP(c, e, True) for c, e in zip(contours, epsilons)]

# Filter out contours that aren't rectangular. Those that aren't rectangular
# are probably noise.
approx_rects = [p for p in approx_polys if len(p) == 4]
bounding_rects = [cv2.boundingRect(a) for a in approx_polys]

# Filter out rectangles that are too narrow or too short.
MIN_RECT_WIDTH = 40
MIN_RECT_HEIGHT = 10
bounding_rects = [
    r for r in bounding_rects if MIN_RECT_WIDTH < r[2] and MIN_RECT_HEIGHT < r[3]
]

# The largest bounding rectangle is assumed to be the entire table.
# Remove it from the list. We don't want to accidentally try to OCR
# the entire table.
largest_rect = max(bounding_rects, key=lambda r: r[2] * r[3])
bounding_rects = [b for b in bounding_rects if b is not largest_rect]

cells = [c for c in bounding_rects]
#+END_SRC

** Sorting the bounding rectangles

We want to process these from left-to-right, top-to-bottom.

I've thought of a straightforward algorithm for it, but it could probably be
made more efficient.

We'll find the most rectangle with the most top-left corner. Then we'll find all
of the rectangles that have a center that is within the top-y and bottom-y
values of that top-left rectangle. Then we'll sort those rectangles by the x
value of their center. We'll remove those rectangles from the list and repeat.

#+NAME: sort-contours
#+BEGIN_SRC python :results none
def cell_in_same_row(c1, c2):
    c1_center = c1[1] + c1[3] - c1[3] / 2
    c2_bottom = c2[1] + c2[3]
    c2_top = c2[1]
    return c2_top < c1_center < c2_bottom

orig_cells = [c for c in cells]
rows = []
while cells:
    first = cells[0]
    rest = cells[1:]
    cells_in_same_row = sorted(
        [
            c for c in rest
            if cell_in_same_row(c, first)
        ],
        key=lambda c: c[0]
    )

    row_cells = sorted([first] + cells_in_same_row, key=lambda c: c[0])
    rows.append(row_cells)
    cells = [
        c for c in rest
        if not cell_in_same_row(c, first)
    ]

# Sort rows by average height of their center.
def avg_height_of_center(row):
    centers = [y + h - h / 2 for x, y, w, h in row]
    return sum(centers) / len(centers)

rows.sort(key=avg_height_of_center)
#+END_SRC

To test if this code works, let's try sorting the bounding rectangles and
numbering them from right to left, top to bottom.

#+HEADER: :post html-image-size(text=*this*, width="500px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
import cv2
image = cv2.imread("resources/examples/example-table.png", cv2.IMREAD_GRAYSCALE)
<<blur>>
<<threshold>>
<<lines-of-table>>
<<bounding-rects>>
<<sort-contours>>

FONT_SCALE = 0.7
FONT_COLOR = (127, 127, 127)
for i, row in enumerate(rows):
    for j, cell in enumerate(row):
        x, y, w, h = cell
        cv2.putText(
            image,
            "{},{}".format(i, j),
            (int(x + w - w / 2), int(y + h - h / 2)),
            cv2.FONT_HERSHEY_SIMPLEX,
            FONT_SCALE,
            FONT_COLOR,
            2,
        )
cv2.imwrite("resources/examples/example-table-cells-numbered.png", image)
#+END_SRC

#+ATTR_HTML: :width 500px :height 100%
[[file:resources/examples/example-table-cells-numbered.png]]

#+NAME: extract-cells-from-table
#+BEGIN_SRC python :noweb yes :eval no
def extract_cell_images_from_table(image):
    <<blur>>
    <<threshold>>
    <<lines-of-table>>
    <<bounding-rects>>
    <<sort-contours>>
    cell_images_rows = []
    for row in rows:
        cell_images_row = []
        for x, y, w, h in row:
            cell_images_row.append(image[y:y+h, x:x+w])
        cell_images_rows.append(cell_images_row)
    return cell_images_rows
#+END_SRC

#+HEADER: :post html-image-size(text=*this*, width="200px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
<<extract-cells-from-table>>
image = cv2.imread("resources/examples/example-table.png", cv2.IMREAD_GRAYSCALE)
cell_images_rows = extract_cell_images_from_table(image)
cv2.imwrite("resources/examples/example-table-cell-1-1.png", cell_images_rows[1][1])
#+END_SRC

#+ATTR_HTML: :width 200px :height 100%
[[file:resources/examples/example-table-cell-1-1.png]]

** Cropping each cell to the text

OCR with Tesseract works best when there is about 10 pixels of white border
around the text.

Our bounding rectangles may have picked up some stray pixels from the horizontal
and vertical lines of the cells in the table. It's probobly just a few pixels,
much fewer than the width of the text. If that's the case, then we can remove
that noise with a simple open morph.

Once the stray border pixels have been removed, we can expand our border using
~copyMakeBorder~.

#+BEGIN_SRC python :eval no :noweb-ref crop-to-text
def crop_to_text(image):
    MAX_COLOR_VAL = 255
    BLOCK_SIZE = 15
    SUBTRACT_FROM_MEAN = -2

    img_bin = cv2.adaptiveThreshold(
        ~image,
        MAX_COLOR_VAL,
        cv2.ADAPTIVE_THRESH_MEAN_C,
        cv2.THRESH_BINARY,
        BLOCK_SIZE,
        SUBTRACT_FROM_MEAN,
    )

    img_h, img_w = image.shape
    horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (int(img_w * 0.5), 1))
    vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, int(img_h * 0.7)))
    horizontal_lines = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN, horizontal_kernel)
    vertical_lines = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN, vertical_kernel)
    both = horizontal_lines + vertical_lines
    cleaned = img_bin - both

    # Get rid of little noise.
    kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
    opened = cv2.morphologyEx(cleaned, cv2.MORPH_OPEN, kernel)
    opened = cv2.dilate(opened, kernel)

    contours, hierarchy = cv2.findContours(opened, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
    bounding_rects = [cv2.boundingRect(c) for c in contours]
    NUM_PX_COMMA = 6
    MIN_CHAR_AREA = 5 * 9
    char_sized_bounding_rects = [(x, y, w, h) for x, y, w, h in bounding_rects if w * h > MIN_CHAR_AREA]
    if char_sized_bounding_rects:
        minx, miny, maxx, maxy = math.inf, math.inf, 0, 0
        for x, y, w, h in char_sized_bounding_rects:
            minx = min(minx, x)
            miny = min(miny, y)
            maxx = max(maxx, x + w)
            maxy = max(maxy, y + h)
        x, y, w, h = minx, miny, maxx - minx, maxy - miny
        cropped = image[y:min(img_h, y+h+NUM_PX_COMMA), x:min(img_w, x+w)]
    else:
        # If we morphed out all of the text, assume an empty image.
        cropped = MAX_COLOR_VAL * np.ones(shape=(20, 100), dtype=np.uint8)
    bordered = cv2.copyMakeBorder(cropped, 5, 5, 5, 5, cv2.BORDER_CONSTANT, None, 255)
    return bordered
#+END_SRC

#+HEADER: :post html-image-size(text=*this*, width="200px")
#+BEGIN_SRC python :noweb no-export :results none :exports both
import cv2
import numpy as np
<<crop-to-text>>
image = cv2.imread("resources/examples/example-table-cell-1-1.png", cv2.IMREAD_GRAYSCALE)
image = crop_to_text(image)
cv2.imwrite("resources/examples/example-table-cell-1-1-cropped.png", image)
#+END_SRC

#+ATTR_HTML: :width 200px :height 100%
[[file:resources/examples/example-table-cell-1-1-cropped.png]]

** OCR each cell

If we cleaned up the images well enough, we might get some accurate OCR!

There is plenty that could have gone wrong along the way.

The first step to troubleshooting is to view the intermediate images and see if
there's something about your image that is obviously abnormal, like some really
thick noise or a wrongly detected table.

If everything looks reasonable but the OCR is doing something like turning a
period into a comma, then you might need to do some custom Tesseract training.

#+BEGIN_SRC python :noweb-ref ocr-image :eval no
def ocr_image(image, config):
    return pytesseract.image_to_string(
        image,
        config=config
    )
#+END_SRC

The second argument passed to ~ocr_image~ is a string of the command line arguments passed directly to ~tesseract~. You can view the available options at [[https://github.com/tesseract-ocr/tesseract/blob/master/doc/tesseract.1.asc#options]]

If no options are passed to ~tesseract~, then language defaults to english. This means ~tesseract~ needs to be able to find a file named ~eng.traineddata~ on whatever path it searches for languages.

This python package comes with ~eng.traineddata~ and ~table-ocr.traineddata~. ~table-ocr.traineddata~ is a personal model that I've found to be more accurate for my use case. You should train your own to maximize accuracy.

When you ~pip install~ this package, the traineddata gets copied to a ~tessdata~ folder in the same directory in which ~pip~ installs the package.

The ~ocr_image~ package in this repo defaults to using the ~--tessdata-dir~ option to the package's ~tessdata~ directory in the package install location and the ~-l~ option to the ~table_ocr~ language.

#+BEGIN_SRC python :noweb no-export :exports both
import pytesseract
import cv2
import numpy as np
import math
image = cv2.imread("resources/examples/example-table-cell-1-1.png", cv2.IMREAD_GRAYSCALE)
<<crop-to-text>>
<<ocr-image>>
image = crop_to_text(image)
ocr_image(image, "--psm 7")
#+END_SRC

#+RESULTS:
: 9.09

* Demo

I wanted to include a demo script that can be used as a quick example.

To run the demo, simply:

1. ~pip3 install table_ocr~
2. ~python3 -m table_ocr.util.url_img_to_csv "https://2ptidz4dnkwy36mu2on9rps1-wpengine.netdna-ssl.com/wp-content/uploads/2015/11/Scanning-Mirror-Data-1.png"~


All of the modules work with filepaths, so whatever you're working with needs to be saved to the fileystem so we can access it by its filename. There is no particular reason for this other than it was the most convenient implementation at the time. This just as well could be modified to accept file-like objects for a lot of the code and we could do all of the work in-memory without storing things to disk.

#+NAME: helper download image to tempdir
#+BEGIN_SRC python
def download_image_to_tempdir(url, filename=None):
    if filename is None:
        filename = os.path.basename(url)
    response = requests.get(url, stream=True)
    tempdir = table_ocr.util.make_tempdir("demo")
    filepath = os.path.join(tempdir, filename)
    with open(filepath, 'wb') as f:
        for chunk in response.iter_content():
            f.write(chunk)
    return filepath
#+END_SRC

This demo starts from an image rather than from a PDF. The concepts should still be apparent. But starting from a PDF would make the demo less demo-able since it would require the person running the demo to have Poppler installed for ~pdftoimages~.

The ~main~ function of ~extract_tables~ takes a list of filepaths of images. It will attempt to find bounding boxes of all tables in the images and return a list of tuples of (<image filepath>, <list of filepaths of found and cropped out tables>)

#+BEGIN_SRC python :tangle table_ocr/demo/__init__.py :mkdirp yes

#+END_SRC

#+NAME: demo main
#+BEGIN_SRC python :tangle table_ocr/demo/__main__.py :mkdirp yes :noweb yes
<<demo imports>>
<<helper download image to tempdir>>

def main(url):
    image_filepath = download_image_to_tempdir(url)
    image_tables = table_ocr.extract_tables.main([image_filepath])
    print("Running `{}`".format(f"extract_tables.main([{image_filepath}])."))
    print("Extracted the following tables from the image:")
    print(image_tables)
    for image, tables in image_tables:
        print(f"Processing tables for {image}.")
        for table in tables:
            print(f"Processing table {table}.")
            cells = table_ocr.extract_cells.main(table)
            ocr = [
                table_ocr.ocr_image.main(cell, None)
                for cell in cells
            ]
            print("Extracted {} cells from {}".format(len(ocr), table))
            print("Cells:")
            for c, o in zip(cells[:3], ocr[:3]):
                with open(o) as ocr_file:
                    # Tesseract puts line feeds at end of text.
                    # Stript it out.
                    text = ocr_file.read().strip()
                    print("{}: {}".format(c, text))
            # If we have more than 3 cells (likely), print an ellipses
            # to show that we are truncating output for the demo.
            if len(cells) > 3:
                print("...")
            return table_ocr.ocr_to_csv.text_files_to_csv(ocr)

if __name__ == "__main__":
    csv_output = main(sys.argv[1])
    print()
    print("Here is the entire CSV output:")
    print()
    print(csv_output)
#+END_SRC

#+NAME: demo imports
#+BEGIN_SRC python
import os
import sys

import requests
import table_ocr.util
import table_ocr.extract_tables
import table_ocr.extract_cells
import table_ocr.ocr_image
import table_ocr.ocr_to_csv
#+END_SRC

* Files
:PROPERTIES:
:header-args: :mkdirp yes :noweb yes
:END:

#+BEGIN_SRC python :tangle table_ocr/__init__.py :mkdirp yes :results none

#+END_SRC

** setup.py
#+BEGIN_SRC python :tangle setup.py :results none
import os
import setuptools

this_dir = os.path.abspath(os.path.dirname(__file__))
with open(os.path.join(this_dir, "README.txt"), encoding="utf-8") as f:
    long_description = f.read()

setuptools.setup(
    name="table_ocr",
    version="0.2.5",
    author="Eric Ihli",
    author_email="eihli@owoga.com",
    description="Extract text from tables in images.",
    long_description=long_description,
    long_description_content_type="text/plain",
    url="https://github.com/eihli/image-table-ocr",
    packages=setuptools.find_packages(),
    package_data={
        "table_ocr": ["tessdata/table-ocr.traineddata", "tessdata/eng.traineddata"]
    },
    classifiers=[
        "Programming Language :: Python :: 3",
        "License :: OSI Approved :: MIT License",
        "Operating System :: OS Independent",
    ],
    install_requires=["pytesseract~=0.3", "opencv-python~=4.2", "numpy~=1.19", "requests>=2"],
    python_requires=">=3.6",
)
#+END_SRC

** table_ocr
*** table_ocr/__init__.py
#+BEGIN_SRC python :tangle table_ocr/__init__.py :results none :exports none

#+END_SRC

*** table_ocr/util.py

#+BEGIN_SRC python :tangle table_ocr/util.py :results none
from contextlib import contextmanager
import functools
import logging
import os
import tempfile

<<get-logger>>

@contextmanager
def working_dir(directory):
    original_working_dir = os.getcwd()
    try:
        os.chdir(directory)
        yield directory
    finally:
        os.chdir(original_working_dir)


def make_tempdir(identifier):
    return tempfile.mkdtemp(prefix="{}_".format(identifier))
#+END_SRC

*** table_ocr/pdf_to_images/
**** table_ocr/pdf_to_images/__init__.py
#+NAME: pdf_to_images/__init__.py
#+HEADER: :tangle table_ocr/pdf_to_images/__init__.py
#+BEGIN_SRC python :results none
import os
import re
import subprocess

from table_ocr.util import get_logger, working_dir

logger = get_logger(__name__)

<<pdf-to-images>>

<<fix-orientation>>
#+END_SRC

**** table_ocr/pdf_to_images/__main__.py

Takes a variable number of pdf files and creates images out of each page of the
file using ~pdfimages~ from Poppler. Images are created in the same directory
that contains the pdf.

Prints each pdf followed by the images extracted from that pdf followed by a
blank line.

#+BEGIN_SRC shell :eval no :exports code
python -m table_ocr.prepare_pdfs /tmp/file1/file1.pdf /tmp/file2/file2.pdf ...
#+END_SRC


#+NAME: pdf_to_images/__main__.py
#+HEADER: :tangle table_ocr/pdf_to_images/__main__.py
#+BEGIN_SRC python
import argparse

from table_ocr.util import working_dir, make_tempdir, get_logger
from table_ocr.pdf_to_images import pdf_to_images, preprocess_img

logger = get_logger(__name__)

parser = argparse.ArgumentParser()
parser.add_argument("files", nargs="+")


def main(files):
    pdf_images = []
    for f in files:
        pdf_images.append((f, pdf_to_images(f)))

    for pdf, images in pdf_images:
        for image in images:
            preprocess_img(image)

    for pdf, images in pdf_images:
        print("{}\n{}\n".format(pdf, "\n".join(images)))


if __name__ == "__main__":
    args = parser.parse_args()
    main(args.files)
#+END_SRC

*** table_ocr/extract_tables/
**** table_ocr/extract_tables/__init__.py

#+NAME: extract_tables/__init__.py
#+HEADER: :tangle table_ocr/extract_tables/__init__.py
#+BEGIN_SRC python
import os
import cv2

<<detect-tables>>

def main(files):
    results = []
    for f in files:
        directory, filename = os.path.split(f)
        image = cv2.imread(f, cv2.IMREAD_GRAYSCALE)
        tables = find_tables(image)
        files = []
        filename_sans_extension = os.path.splitext(filename)[0]
        if tables:
            os.makedirs(os.path.join(directory, filename_sans_extension), exist_ok=True)
        for i, table in enumerate(tables):
            table_filename = "table-{:03d}.png".format(i)
            table_filepath = os.path.join(
                directory, filename_sans_extension, table_filename
            )
            files.append(table_filepath)
            cv2.imwrite(table_filepath, table)
        if tables:
            results.append((f, files))
    # Results is [[<input image>, [<images of detected tables>]]]
    return results
#+END_SRC

**** table_ocr/extract_tables/__main__.py

Takes 1 or more image paths as arguments.

Images are opened and read with OpenCV.

Tables are detected and extracted to a new subdirectory of the given image. The
subdirectory will be the filename sans the extension. The tables inside that
directory will be named ~table-000.png~.

If you want to do something with the output, like pipe the paths of the
extracted tables into some other utility, here is a description of the output.

For each image path given as an agument, outputs:

1. The given image path
2. Paths of extracted tables; seperated by newlines
3. Empty newline

#+NAME: extract_tables/__main__.py
#+BEGIN_SRC python :tangle table_ocr/extract_tables/__main__.py :results none
import argparse

from table_ocr.extract_tables import main

parser = argparse.ArgumentParser()
parser.add_argument("files", nargs="+")
args = parser.parse_args()
files = args.files
results = main(files)
for image, tables in results:
    print("\n".join(tables))
#+END_SRC

*** table_ocr/extract_cells/

**** table_ocr/extract_cells/__init__.py

#+BEGIN_SRC python :tangle table_ocr/extract_cells/__init__.py
import cv2
import os

<<extract-cells-from-table>>

def main(f):
    results = []
    directory, filename = os.path.split(f)
    table = cv2.imread(f, cv2.IMREAD_GRAYSCALE)
    rows = extract_cell_images_from_table(table)
    cell_img_dir = os.path.join(directory, "cells")
    os.makedirs(cell_img_dir, exist_ok=True)
    paths = []
    for i, row in enumerate(rows):
        for j, cell in enumerate(row):
            cell_filename = "{:03d}-{:03d}.png".format(i, j)
            path = os.path.join(cell_img_dir, cell_filename)
            cv2.imwrite(path, cell)
            paths.append(path)
    return paths
#+END_SRC

**** table_ocr/extract_cells/__main__.py

Takes as a command line argument a path to an image of a table.

Detects cells in the table and extracts each cell to an image file in a new
~/cells/~ subdirectory in the same directory of the given image's path.

Each cell filename is suffixed with ~<row>-<column>~ so that the filenames can
be sorted lexicographically and will align with reading the cells from
left-to-right, top-to-bottom.

Prints to stdout the lexicographically sorted list of filenames of the extracted
cells.

#+BEGIN_SRC python :tangle table_ocr/extract_cells/__main__.py :results none
import sys

from table_ocr.extract_cells import main

paths = main(sys.argv[1])
print("\n".join(paths))
#+END_SRC

*** table_ocr/ocr_image/
**** table_ocr/ocr_image/__init__.py
#+BEGIN_SRC python :tangle table_ocr/ocr_image/__init__.py
import math
import os
import sys

import cv2
import numpy as np
import pytesseract

def main(image_file, tess_args):
    """
    OCR the image and output the text to a file with an extension that is ready
    to be used in Tesseract training (.gt.txt).

    Tries to crop the image so that only the relevant text gets passed to Tesseract.

    Returns the name of the text file that contains the text.
    """
    directory, filename = os.path.split(image_file)
    filename_sans_ext, ext = os.path.splitext(filename)
    image = cv2.imread(image_file, cv2.IMREAD_GRAYSCALE)
    cropped = crop_to_text(image)
    ocr_data_dir = os.path.join(directory, "ocr_data")
    os.makedirs(ocr_data_dir, exist_ok=True)
    out_imagepath = os.path.join(ocr_data_dir, filename)
    out_txtpath = os.path.join(ocr_data_dir, "{}.gt.txt".format(filename_sans_ext))
    cv2.imwrite(out_imagepath, cropped)
    if not tess_args:
        d = os.path.dirname(sys.modules["table_ocr"].__file__)
        tessdata_dir = os.path.join(d, "tessdata")
        tess_args = ["--psm", "7", "-l", "table-ocr", "--tessdata-dir", tessdata_dir]
    txt = ocr_image(cropped, " ".join(tess_args))
    with open(out_txtpath, "w") as txt_file:
        txt_file.write(txt)
    return out_txtpath

<<crop-to-text>>
<<ocr-image>>
#+END_SRC
**** table_ocr/ocr_image/__main__.py

This does a little bit of cleanup before sending it through tesseract.

Creates images and text files that can be used for training tesseract. See
https://github.com/tesseract-ocr/tesstrain.

#+BEGIN_SRC python :tangle table_ocr/ocr_image/__main__.py :mkdirp yes :results none
import argparse

from table_ocr.ocr_image import main

description="""Takes a single argument that is the image to OCR.
Remaining arguments are passed directly to Tesseract.

Attempts to make OCR more accurate by performing some modifications on the image.
Saves the modified image and the OCR text in an `ocr_data` directory.
Filenames are of the format for training with tesstrain."""
parser = argparse.ArgumentParser(description=description)
parser.add_argument("image", help="filepath of image to perform OCR")

args, tess_args = parser.parse_known_args()
print(main(args.image, tess_args))
#+END_SRC
*** table_ocr/ocr_to_csv/
**** table_ocr/ocr_to_csv/__init__.py
#+BEGIN_SRC python :tangle table_ocr/ocr_to_csv/__init__.py
import csv
import io
import os


def text_files_to_csv(files):
    """Files must be sorted lexicographically
    Filenames must be <row>-<colum>.txt.
    000-000.txt
    000-001.txt
    001-000.txt
    etc...
    """
    rows = []
    for f in files:
        directory, filename = os.path.split(f)
        with open(f) as of:
            txt = of.read().strip()
        row, column = map(int, filename.split(".")[0].split("-"))
        if row == len(rows):
            rows.append([])
        rows[row].append(txt)

    csv_file = io.StringIO()
    writer = csv.writer(csv_file)
    writer.writerows(rows)
    return csv_file.getvalue()

def main(files):
    return text_files_to_csv(files)
#+END_SRC
**** table_ocr/ocr_to_csv/__main__.py

#+BEGIN_SRC python :tangle table_ocr/ocr_to_csv/__main__.py
import argparse
import os

from table_ocr.ocr_to_csv import text_files_to_csv

parser = argparse.ArgumentParser()
parser.add_argument("files", nargs="+")


def main(files):
    print(text_files_to_csv(files))


if __name__ == "__main__":
    args = parser.parse_args()
    files = args.files
    files.sort()
    main(files)
#+END_SRC

* Utils

The following code lets us specify a size for images when they are exported to
html.

Org supports specifying an export size for an image by putting the ~#+ATTR_HTML:
:width 100px~ before the image. But since our images are in a results drawer, we
need a way for our results drawer to do that for us automatically.

Adding ~#+ATTR_HTML~ after the beginning of the result block introduces a new
problem. Org-babel no longer recognizes the result as a result block and doesn't
remove it when a src block is re-evaluated, so we end up just appending new
results on each evaluation.

There is nothing configurable that will tell org-babel to remove our line. But
we can define a function to do some cleanup and then add it as a before hook
with ~advice-add~.

#+NAME: html-image-size
#+BEGIN_SRC emacs-lisp :var text="" :var width="100%" :var height="100%" :results raw :export code
(concat "#+ATTR_HTML: :width " width " :height " height "\n[[file:" text "]]")
#+END_SRC

#+BEGIN_SRC emacs-lisp :results none
(defun remove-attributes-from-src-block-result (&rest args)
  (let ((location (org-babel-where-is-src-block-result))
        (attr-regexp "[ 	]*#\\+ATTR.*$"))
    (when location
      (save-excursion
        (goto-char location)
        (when (looking-at (concat org-babel-result-regexp ".*$"))
          (next-line)
          (while (looking-at attr-regexp)
            (kill-whole-line)))))))

(advice-add 'org-babel-remove-result :before #'remove-attributes-from-src-block-result)
(advice-add 'org-babel-execute-src-block :before #'remove-attributes-from-src-block-result)
#+END_SRC

** Logging

#+BEGIN_SRC python :eval query :noweb-ref get-logger
def get_logger(name):
    logger = logging.getLogger(name)
    lvl = os.environ.get("PY_LOG_LVL", "info").upper()
    handler = logging.StreamHandler()
    formatter = logging.Formatter(logging.BASIC_FORMAT)
    handler.setFormatter(formatter)
    logger.addHandler(handler)
    handler.setLevel(lvl)
    logger.setLevel(lvl)
    return logger
#+END_SRC