To provide the best results of our plant identification API, our tech stack uses Google’s Tensorflow and a python-based heavy lifting. Minor part of the service is based on Google Cloud and Microsoft Azure.
The neural network’s architecture is based on the Inception model, but we have implemented a bunch of customizations, for example to estimate identification certainty, select best adaptive representative photos and improve the accuracy with meta attributes like GPS coordinates.
Our newest model can identify 10 675 plant species. The set covers the most commonly requested plants from the United States, European Union, India, and Australia.
We are continually testing model accuracy. The average success rate ranges between 80%–90%, although it strongly depends on photo and plant quality. Our goal is to improve the average over 85% while extending the spectrum of identified species.