Image recognition done differently.

Case study

Image recognition done differently.

Implementing a solution for extracting detailed data from scanned invoices.

Challenge

The Telco company was dedicating a lot of man-power on entering the data from invoices into their system. They had a semi-automatic solution but results delivered from it needed lots of corrections and caused problems down the line.

Solutions

We have started by analysing the process of invoice processing to understand the volume and distribution of various types of documents. Then we understood that we should focus on the most common vendors to make sure that we make a fast impact and collect feedback. Having well defined targets allowed us to use lightweight methods which led to great results without using powerful but overly costly for this case methods such as image processing using deep learning.

Technology & Tools

Python

Scala

Tesseract

Client

A large enterprise leader in the telecommunications industry across the Nordic-Baltic region. Aiming for automating invoice processing and limiting manual labour.

Opportunity

Leveraging automatic invoice parsing has a potential to speed up the processing. It also means significant cost reductions for any large company. The Telco client was aware of the opportunity and had already utilised a tool meant for this purpose, however the tools did not perform as expected. Incorrect data extraction results provided by the tool led to errors. These were often found later in the process which caused friction and a long manual process of correcting them. Quick support from specialists was needed to re-architect and extend the solution for the perfect fit and delivery.

Delivery

The implementation was preceded with careful analysis of what the process looked like, what were the main pain points. We identified that some types of invoices are significantly more common than the other ones. Other than that not all invoice data was equally important. Bearing that in mind we focused on the most critical fields of most common invoices so the customer can get immediate impact. Even though deep learning is quite advanced in the image processing area and it was tempting to use it for tackling this challenge, we understood that applying it to this problem would be time consuming and complex. During the analysis period we understood that the business would not benefit from very generic and complex solutions. By precisely defining the focus area we managed to use simplified methods like using already existing open source tools and pattern matching.