Create Your Own Working Pokédex!

Ever wanted your own Pokédex that detects Pokémon in the wild? Adrian Rosebrock’s might have the answer.

Using deep learning, Adrian Rosebrock is turned a Raspberry Pi into a Pokédex. He trained a neural network to identify Pokémon using a dataset of 1191 Pokémon images. And its accuracy? 96.84%. The build itself uses a Raspberry Pi 3 with camera module attached to a Raspberry Pi 7″ touch display. I’m sure with a bit of ingenuity, someone could 3D print a red chassis but one step at a time.

Adrian has outlined the model’s limitations including colour influencing classifications. This means it’s more likely to be wrong if multiple Pokémon have similar colours. He advises to have “at least 500-1,000 images per class when training a Convolutional Neural Network”. Just go on DeviantArt, you’ll find loads of Pokémon images. Just be careful which ones you choose.

Check out a project video below and read more about it on Adrian’s blog.

(via Raspberry Pi)