IE was engaged to assist one of Australia's largest logistics companies to test the feasibility of building a prototype that scans and matches handwritten addresses in their internal database. The solution needed to return three possible matched addresses within four seconds via API calls and have a performance accuracy of at least 70%.
How might we test the feasibility of building a machine learning system that reads handwritten addresses and returns matches within a given database.
Machine learning (ML) is the use of computer systems that learn and adapt without following explicit instructions, by using algorithms and statistical models to analyse and draw inferences from patterns in data.
Optical Character Recognition (OCR) is the process of converting images containing text (usually printed but in this case handwritten) into machine-encoded text. OCR of handwriting is still a relatively novel and unsolved problem in ML applications.
The final product was a mobile app that captured the image of the parcel address (our input), requested API calls (our process), and displayed three likely addresses for the user to select (our output). At the start of the project, our original goal was to achieve a match rate of 70%. By the end of the engagement, the team was able to achieve a performance baseline match rate of an incredible 77%, in which one of the three returned addresses was correct. This rose to 81% (adjusted match rate) when filtering out the addresses that even humans could not find in the database.
All of this was achieved in 4 weeks.