Instantly share code, notes, and snippets.

Embed
What would you like to do?
Using Python 3.x and Google Cloud Vision API to OCR scanned documents to extract structured data

Using Python 3 + Google Cloud Vision API's OCR to extract text from photos and scanned documents

Just a quickie test in Python 3 (using Requests) to see if Google Cloud Vision can be used to effectively OCR a scanned data table and preserve its structure, in the way that products such as ABBYY FineReader can OCR an image and provide Excel-ready output.

The short answer: No. While Cloud Vision provides bounding polygon coordinates in its output, it doesn't provide it at the word or region level, which would be needed to then calculate the data delimiters.

On the other hand, the OCR quality is pretty good, if you just need to identify text anywhere in an image, without regards to its physical coordinates. I've included two examples:

####### 1. A low-resolution photo of road signs

Courtesy of SPUI on Wikicommons:

Jughandle_signage

1a. Google GMail CAPTCHA (circa 2009)

Added this out of curiousity: a sample taken from Google's 2009 research paper, What’s Up CAPTCHA?; A CAPTCHA Based On Image Orientation

captcha

####### 2. An image (PDF to PNG) of a spreadsheet

Courtesy of Eli Lilly:

image

You can read more about getting started with the Google Cloud Vision API in its official docs. My Python script is a somewhat simplified version of the official instructions here:

https://github.com/GoogleCloudPlatform/cloud-vision/tree/master/python/text

You first have to set up a Google developer account and get an API key (the API allows 1000 free requests a month):

https://cloud.google.com/vision/docs/auth-template/cloud-api-auth#set_up_an_api_key

How to run

The cloudvisreq.py script is included at the bottom of this gist.

$  python cloudvisreq.py API_KEY image1.jpg image2.png

Results

Road signs

    Bounding Polygon:
{'vertices': [{'x': 16, 'y': 21}, {'x': 772, 'y': 21}, {'x': 772, 'y': 322}, {'x': 16, 'y': 322}]}
    Text:
WARRENVILLE RD
NORTH
SOUTH
ALL TURNS
FROM
WASHINGTON AVE
RIGHT LANE
GREEN BROOK
U AND LEFT
DUNELLEN
TURNS
ALL TURNS f
A

GMail Captcha

Sorry spammers, you probably won't get far using Google's API against its old CAPTCHA system (nevermind its current one):

    Bounding Polygon:
{'vertices': [{'y': 73, 'x': 142}, {'y': 73, 'x': 339}, {'y': 173, 'x': 339}, {'y': 173, 'x': 142}]}
    Text:
ngly:h

Spreadsheet

    Bounding Polygon:
{'vertices': [{'y': 272, 'x': 212}, {'y': 272, 'x': 3066}, {'y': 2295, 'x': 3066}, {'y': 2295, 'x': 212}]}
    Text:
Lilly other Health Care Professional Registry
Data updated on Monday, March 3, 2014
Payments Made: Q1-Q4 2013
The Other Health Care Professional Registry reports direct and indirect payments by Lilly, as well as Lilly's portion of alliance partnership paymen
o health care professionals other than physicians serving as faculty
members. When the "Entity Paid
a company, hospital, or university, and there
a different name on the provider of service
reflects an indirect payment which may or may not actually have been received in
whole or in part by the
ted service provider. Copyright 2014 Eli Lilly and Company. All rights reserved
Note: Due to differences in scope and definitions, the data reported in this report may differ from data included in reports submitted by Lilly for compliance with state payment reporting laws
Payments
All Amounts in US Dollars
Provider of Service
*Payments for reimbursed expenses are not compensation
Entity Paid
Patient Education Professional
Education
Education
WARD, JANET
OH ABDALLAH, RITA $14,025 $500 $2,823 S17348
ABDALLAH, RITA
FAIRVIEW PARK
OH AGOSTI, CAROL $1,050 1,631 $355 $3036
AGOSTI, CAROL
TOLEDO
AINSWORTH, ABBY CORPUS CHRISTI AINSWORTH, ABBY $113 $113
AKERS, REBECCA VINCENNES IN AKERS, REBECCA
$2,400 $2,400
IALCHEMIPHARMA LLC, WAYNE PA BELAZI, DEA $11,889 $1.99 $12,088
MEADOW VISTA
ALEMAN, MARY
ALEMAN, MARY
ALEXANDER, LISA
ALEXANDER, LISA
FALLEN, CONNETTE COTTAGE GROVE
MN ALLEN, CONNETTE $6,113 $875 $2,347 $9,335
MALLISON HEINRICH, LL. LUBBOCK HEINRICH, ALLISON $4,200 $1.238 $5.438
ALLISON, CRYSTAL GOODYEAR
AZ ALLISON, CRYSTAL s250 $250
MALLISON, SUSAN
EASTON
PA ALLISON, SUSAN $7950 $2.138 $2.252 $12.340
ALVAREZ, MICHELLE MCALLEN TX ALVAREZ, MICHELLE
S4,650 $750 $244 $5,644
GRANADA HILLS CA ANGELES, ADA
$39,150 $11,456 $938 $9,181 $60,725
ANGELES WORLD, INC.,
ARBOLEDA, JANE
MIAMI FL ARBOLEDA, JANE
S375 $4,875 $549 $5,799
WACO ARMSTRONG, JULIE
$2,400 $750 $3,150
ARMSTRONG, JULIE
S900 $2,006 $657 $3,563
ARNOLD, MARY BETH
AUGUSTA
GA ARNOLD, MARY BETH
ARRAMBIDE, ROBIN KATY TX ARRAMBIDE, ROBIN
$11719 $11719
ARRINGTON, WANDA CHARLOTTE NC ARRINGTON, WANDA
S2,531 $258 $2789
PA ASH FORD, RICHARD
$3,525 S339 $3,864
WILKES BARRE
ASHFORD, RICHARD
ATTANASIO, MICHAEL SEWELL NJ ATTANASIO, MICHAEL
$450 $450
INDIANA PA AvOLIO JOHN $225 $225
AVOLIO JOHN
BABEY, CHRISTINE SCOTTSDALE AZ BABEY, CHRISTINE $5.100 $13,294 $2.135 $20,529
DALLAS BAILEY-GRAY, PATRICK 226 $226
BAILEY-GRAY, PATRICK
DECATUR IL BAKER, BENITA $17.100 S11738 $6,884 $35,722
BAKER, BENITA

Tesseract comparison

Tesseract (version 3.04 as of Feb. 2016), the open-source OCR tool currently maintained by Google, doesn't perform meaningfully in the photo of the road signs, but it is generally pretty strong in the case of screenshotted text. For comparison's sake, here is its output of the tabular data example -- unlike the Cloud Vision API, it does attempt to preserve some of the line layout -- and it has the option of providing HOCR output which can then be used to further define spatial layout.

(note: I've preserved the whitespace in the output, so there are a few dozen blank lines at the bottom...keep scrolling to get to the next section...)

 

o Lilly Other Health Care Professional Registry
Payments Made: Q1-Q4 2013 Data updated on Monday, March 3, 2014

 

The Other Health Care Professional Registry reports direct and indirect payments by Lilly, as well as Lilly's portion of alliance partnership payments to health care professionals other than physicians serving as faculty
members. When the "Entity Paid" is a company, hospital, or university, and there is a different name on the provider of service, it reflects an indirect payment which may or may not actually have been received in
whole or in part by the listed service provider. Copyright © 2014 Eli Lilly and Company. All rights reserved.

Note: Due to differences in scope and definitions, the data reported in this report may differ from data included in reports submitted by Lilly for compliance with state payment reporting laws.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Payments*
(All Amounts in US Dollars)
Provider of Service *Payments for re u bursed expenses are not compensatio
Advising]
H
Patient Education “2:13:33 consulting & Certain
Name Location State Name . International Travel-Related 2013 Total
Programs Education .
Programs Education Expenses
Programs
A1 CERTIFIED DIABETES EDUCATORS,
LLC GILBERT AZ WARD, JAN ET $300 $1,631 $129 $2,060
ABDALLAH, RITA FAIRVIEW PARK OH ABDALLAH, RITA $14,025 $500 $2,823 $17,348
ADAMS, MARY ELLEN CINCINNATI OH ADAMS, MARY ELLEN $3,450 $1,594 $5,044
AGOSTI, CAROL TOLEDO OH AGOSTI, CAROL $1,050 $1,631 $355 $3,036
AINSWORTH, ABBY CORPUS CHRISTI TX AINSWORTH, ABBY $113 $113
AKERS, REBECCA VINCENNES IN AKERS, REBECCA $2,400 $2,400
ALCHEMIPHARMA LLC, WAYNE PA BELAZI, DEA $11,889 $199 $12,088
ALEMAN, MARY MEADOW VISTA CA ALEMAN, MARY $3,300 $1,294 $346 $4,940
ALEXANDER, LISA EVANSVILLE IN ALEXANDER, LISA $3,000 $1,988 $1,077 $6,065
ALLEN, CONNETI'E COTTAGE GROVE MN ALLEN, CONNETI'E $6,113 $875 $2,347 $9,335
ALLISON HEINRICH, L.L. LUBBOCK TX HEINRICH, ALLISON $4,200 $1,238 $5,438
ALLISON, CRYSTAL GOODYEAR AZ ALLISON, CRYSTAL $250 $250
ALLISON, SUSAN EASTON PA ALLISON, SUSAN $7,950 $2,138 $2,252 $12,340
ALVAREZ, MICHELLE MCALLEN TX ALVAREZ, MICHELLE $4,650 $750 $244 $5,644
ANDARIESE, JUDITH BROOKLYN NY AN DARIESE, JUDITH $12,450 $338 $1,029 $13,817
ANGELES WORLD, INC., GRANADA HILLS CA ANGELES, ADA $39,150 $11,456 $938 $9,181 $60,725
ARBOLEDA, JANE MIAMI FL ARBOLEDA, JANE $375 $4,875 $549 $5,799
ARMSTRONG, JULIE WACO TX ARMSTRONG, JULIE $2,400 $750 $3,150
ARNOLD, MARY BETH AUGUSTA GA ARNOLD, MARY BETH $900 $2,006 $657 $3,563
ARRAMBIDE, ROBIN KATY TX ARRAMBIDE, ROBIN $1,719 $1,719
ARRINGTON, WANDA CHARLOTTE NC ARRINGTON, WANDA $2,531 $258 $2,789
ASHFORD, RICHARD WILKES BARRE PA ASHFORD, RICHARD $3,525 $339 $3,864
ATTANASIO, MICHAEL SEWELL NJ AlTANASIO, MICHAEL $450 $450
AVOLIO, JOHN INDIANA PA AVOLIO, JOHN $225 $225
BABEY, CHRISTINE SCOTTSDALE AZ BABEY, CHRISTINE $5,100 $13,294 $2,135 $20,529
BAILEY-G RAY, PATRICK DALLAS TX BAILEY-G RAY, PATRICK $226 $226
BAIRD, DENISE LANCASTER PA BAIRD, DENISE $3,000 $638 $176 $3,814
BAKER, BENITA DECATUR IL BAKER, BENITA $17,100 $11,738 $6,884 $35,722

 

 

Performance and latency

I didn't rigorously test this so these are just rough averages/medians of how long it took for the entire script (including any network latency) to complete:

  • Road signs: 2.1 seconds
  • Spreadsheet: 6.8 seconds
  • Spreadsheet (using Tesseract): 4.1 seconds

Cloud Vision probably isn't intended for picking apart text documents. Occasionally, the API would fail on the spreadsheet image with this result:

{
  "error": {
    "code": 4,
    "message": "image-annotator::RPC deadline exceeded.: Backend timeout!"
  }
}

For a more robust experience, you probably want to follow the example linked from the API's official docs:

https://github.com/GoogleCloudPlatform/cloud-vision/tree/master/python/text

It includes using the googleapiclient library, which has various conveniences including a num_retries argument.

Authenticating via oauth2 JSON credentials

If you want to use the googleapiclient, which includes authenticating via oauth2 credentials, here's a variation of how to authenticate a service request as shown in the official docs, except from a given filename (i.e. as opposed to setting an environment variable and calling GoogleCredentials.get_application_default()):

from googleapiclient import discovery
from oauth2client.client import GoogleCredentials
DISCOVERY_URL = 'https://{api}.googleapis.com/$discovery/rest?version={apiVersion}'
def get_vision(oauth2_creds_filename, service_url=DISCOVERY_URL):
    """
    Read oauth2 credentials and return a Google service object,
      which you can then invoke like this:

    ("vision" is the service object)
    request = vision.images().annotate(body={'requests': img_requests_data})
    vision_response_dict = request.execute(num_retries=5)

    """
    creds = GoogleCredentials.from_stream(oauth2_creds_filename)
    service = discovery.build('vision', 'v1', credentials=creds,
                              discoveryServiceUrl=DISCOVERY_URL)
    return service
from base64 import b64encode
from os import makedirs
from os.path import join, basename
from sys import argv
import json
import requests
ENDPOINT_URL = 'https://vision.googleapis.com/v1/images:annotate'
RESULTS_DIR = 'jsons'
makedirs(RESULTS_DIR, exist_ok=True)
def make_image_data_list(image_filenames):
"""
image_filenames is a list of filename strings
Returns a list of dicts formatted as the Vision API
needs them to be
"""
img_requests = []
for imgname in image_filenames:
with open(imgname, 'rb') as f:
ctxt = b64encode(f.read()).decode()
img_requests.append({
'image': {'content': ctxt},
'features': [{
'type': 'TEXT_DETECTION',
'maxResults': 1
}]
})
return img_requests
def make_image_data(image_filenames):
"""Returns the image data lists as bytes"""
imgdict = make_image_data_list(image_filenames)
return json.dumps({"requests": imgdict }).encode()
def request_ocr(api_key, image_filenames):
response = requests.post(ENDPOINT_URL,
data=make_image_data(image_filenames),
params={'key': api_key},
headers={'Content-Type': 'application/json'})
return response
if __name__ == '__main__':
api_key, *image_filenames = argv[1:]
if not api_key or not image_filenames:
print("""
Please supply an api key, then one or more image filenames
$ python cloudvisreq.py api_key image1.jpg image2.png""")
else:
response = request_ocr(api_key, image_filenames)
if response.status_code != 200 or response.json().get('error'):
print(response.text)
else:
for idx, resp in enumerate(response.json()['responses']):
# save to JSON file
imgname = image_filenames[idx]
jpath = join(RESULTS_DIR, basename(imgname) + '.json')
with open(jpath, 'w') as f:
datatxt = json.dumps(resp, indent=2)
print("Wrote", len(datatxt), "bytes to", jpath)
f.write(datatxt)
# print the plaintext to screen for convenience
print("---------------------------------------------")
t = resp['textAnnotations'][0]
print(" Bounding Polygon:")
print(t['boundingPoly'])
print(" Text:")
print(t['description'])
@sabuncu

This comment has been minimized.

sabuncu commented Mar 27, 2016

Thank you! This is so useful, and one of a handful of interesting work currently available for Google Cloud Vision.

@lf94

This comment has been minimized.

lf94 commented Sep 3, 2016

Nice write up. I'm trying to simulate what Google Books does with scanning books, because I have a bunch of pdfs. I was wondering if Google's Cloud Vision would be good for this, but apparently not. I'll try tesseract but I heard it isn't great for preserving page layout? Then again, OCRopus (which is what Google Books uses apparently) doesn't seem to do it either (based on my short amount of searching...).

@Milap7

This comment has been minimized.

Milap7 commented Sep 26, 2016

Hello, I am having trouble running this file. I got a couple of errors "Invalid Syntax for '*image' line 46" and " makedirs() got an unexpected keyword argument 'exist_ok'. I am new to python so this is a bit confusing for me. Also the api_key link is a dead link.

@Anid4u2c

This comment has been minimized.

Anid4u2c commented Nov 23, 2016

This is great preliminary research. I'm thinking a less complex table would resolve the "Backend timeout!" error. A low-level test with 6 columns, 7 rows, and 1 header row with wrapped text provided similar (-/+ 5 px) coordinates for the beginning (top-left & bottom-left) of the column-based text results. The position of the header row text was not as patterned, because the text was centered prior to the image's creation - the bounding coordinates for the first column and last column could hypothetically create a frame, providing that the header row's height - which could be increased by a subsequent column's text (if wrapped) is higher or lower than the previous bounding coordinates. My hope is that small-sample-group statistical analysis would predict the position of the text within each cell, for each column and row - by grouping their bounds. Those bounds would then be used to identify the grid required to rebuild the table. Further interest motivates me to implement machine learning to assist in the identification of tables, rather than parsing an image such as the "road signs" example above, and identifying it as a table.

This is the image I analyzed:
image295

Anyone can "Try the API" here: https://cloud.google.com/vision/

@midida

This comment has been minimized.

midida commented Apr 18, 2017

@Milap7:
I think you need to use python3 instead of python2. I got the same error with python2.

@rahpalrah

This comment has been minimized.

rahpalrah commented Jul 2, 2018

How do i specify any other language from English to French or Hindi. These are language supported by Google Vision API but unable to figure out where to add for the extraction

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment