Skip to content

Latest commit

 

History

History
163 lines (130 loc) · 6.75 KB

Readme.md

File metadata and controls

163 lines (130 loc) · 6.75 KB
Raw

tinvois-parser

An API to extract information from images of invoices/receipts. It extracts date, total amount, amount excluding VAT and the merchant name.

Try it here: https://tinvois-parser.azurewebsites.net. (http://tinvois-parser.azurewebsites.net shows the swagger UI). The authorization token is "github_users". See example jupyter notebook for sample code how to call.

Why I made it?

We wanted to develop an app to help freelancers in tax declaration, specifically organizing receipts. It is available here: https://tinvois.de. We wanted to keep extra costs including 3rd party tools and marketing really low to be able to keep the price of the app small.

One steps is to extract information from receipts photos. The open source solutions were not good enough. There are commercial API's which either do not work or are too expensive or both. So I developed it myself.

Considering that Google charges only 1.5$ for OCRing each 1K images, even gives 1K calls per month for free, it is almost free. So let's make it really good.

How it works

In the following steps

  • Sending the image to google Cloud Vision API to extract text
  • Putting the results in a pandas DataFrame (I am a data scientist, I love DataFrames :))
  • Preprocess the results as follows
    • Lower case all the strings
    • Join some words that we know come after each other and are meaningful together
    • In case the number, say, "12. 13" is detected as two words, join them
    • Tokenize the words in "SUM", "NETTO" (amount excluding VAT), "BRUTTO" (gross amount) "VAT", "VALUE" (strings specifying a value) and "OTHER"
    • Convert the values to float and keep them in another DataFrame
    • Detect the rotation and rotate the coordinates back to vertical
    • For each value, extract a feature set. Namely, tokens that appear in front of it, behind it, or on top of it.
  • Extracting information as follows

    • Date: get the date string via string matching. I just look for the first match I find.

    • Total amount and amount excluding vat: Writing rules using features extracted in the preprocessing step. Why not machine learning? Because I did not have data to train. So I had to play the role of ML algorithm based on my knowledge of receipts structure. The rules are set based on receipts common in Germany. Feel free to suggest rules for other countries

    • Merchant name: I listed the most common merchants in Germany. It first tries to string match one from that list. If none of them found, it uses simply the first line in the image. Turns out it works fine.

Sample result

Input image

Sample receipt

Result

{
    "data": {
        "rotation": 90,
        "amount": 225,
        "amountexvat": 205,
        "merchant_name": "Penny",
        "date": "2020-12-04T00:00:00",
        "hash": "0000000000000000071f7fffffffffffffbfffbefffefffbffff001f0000f800"
    }
}

How to use

You can start the API locally either via python or docker desktop.

Common steps

Using python

  • Put the google_auth.json in app/google_auth folder
  • Install requirements
  • Put an environmental variable called SERVER_TO_SERVER_TOKEN in your system containing a custom string. It will be the authorization token for calling the API
  • In terminal navigate to app folder and run python .\manage-local.py

Using docker

There are two options

Option 1 (won't work in windows as mounting folder is not possible)

  • Put the google_auth.json in a folder
  • Run this commend
docker run --name tinvois-parser -d \
    -p 5001:5001 \
    -v <path to folder containing google_auth.json>:/app/google_auth \
    -e SERVER_TO_SERVER_TOKEN=<some string which will be used as authorization token> \
    -e BIND=0.0.0.0:5001 \
    -e MODULE_NAME=manage \
    -e WEB_CONCURRENCY=2 \
    srhumir/tinvois-parser:latest

Option 2

  • Base64 encode the content of google_auth.json
  • Run this command (you might need to remove "\"'s and put the whole command in one line in Windows)
docker run --name tinvois-parser -d \
    -p 5001:5001 \
    -e GOOGLE_AUTH=<the base64 encoded string you get above> \
    -e SERVER_TO_SERVER_TOKEN=<some string which will be used as authorization token> \
    -e BIND=0.0.0.0:5001 \
    -e MODULE_NAME=manage \
    -e WEB_CONCURRENCY=2 \
    srhumir/tinvois-parser:latest

It will pull the image from docker hub and run it. The "latest" tag always corresponds to the latest commit in master branch of this repository.

The API is accessible in localhost:5001. Enter it in your browser to see the swagger UI

What else it can do

  • I added endpoints for detecting edges of paper in the image and also making bird view of the document using the edges. See the example jupyter notebook for an example.

Acknowledgements

TODO's (not necessarily comprehensive)

  • Guess the category of the receipt (grocery, gas, travel etc.)
  • Guess the payment method of the receipt
  • Extend the tests to proper unit tests
  • Run tests using github actions on commit
  • Make it to be able to use Azure OCR API
  • Extract merchant address (maybe using this approach https://doi.org/10.1145/2494188.2494193 available for download ftp://www.kom.tu-darmstadt.de/papers/SMRS13-1.pdf)
  • Implement a small WebUI
  • Produce data for training ML algorithm
  • Do proper image hashing
  • Add some python code for testing the API
  • Improve how it gets the google json file so mounting a folder in the docker command is not necessary
  • Prepare a runnable windows PowerShell docker command
  • Optionally return automatically edited image in parse endpoint
  • Deploy the API in a (free) server so that people can test it