We helped our customer to improve their solution for reading text in images and went from 4% error rate to a much better 0.2%.
Our customer is a Swedish manufacturer, and just like most other similar companies, they work extensively with automation in their production. At one stage, they utilized OCR to read text containing instructions from images. However, the current solution had an error rate of around 3%, which wasn't good enough.
Since we didn't have access to the internet, the only option was to deploy the algorithm directly on the hardware in the factories. Unfortunately, those computers came with CPU only, which is far from optimal for deep learning algorithms. Also, we needed to provide solutions for two different environments. One new based on C# and an older version using VB6.
We managed to implement a deep learning algorithm in both C# and VB6. A vital feature of this implementation is that the algorithm asks factory workers for help whenever it feels uncertain. When that happens, we collect the correct answer and store that image for future training.
When we started the project, the current OCR solution had an error rate of 2-4%. We managed to decrease that to 0.2% which saves the company thousands of hours.
Together with Infranord and Trafikverket, we’ve developed machine learning algorithms for detecting damages on railroad infrastructure.
We helped a large consulting company to automatically detect requirements in their tender process
Using new advancements in connecting text and images, we helped a large media production company implement effective video search without the need for meta-data.