Identify Pokémon Using Deep Learning on a Raspberry Pi

Adrian Rosebrock was, and still is, a fan of Pokémon. This Japanese cartoon sprung onto the scene in 1995, and seems to have never left our collective conciousness via various toys and video game iterations. One interesting item used in this fictional world is called a Pokédex, which allows would-be trainers to identify Pokémon in the wild using some sort of camera and screen interface.

While firmly in the realm of science fiction in 1995, hardware-wise this sounds like a smartphone… or perhaps even a Raspberry Pi with a Camera Module. This advancement of technology wasn’t lost on Rosebrock, who actually built a Pokédex using a Raspberry Pi 3, along with a camera module, touchscreen, and a battery to make everything portable. Using Keras and TensorFlow software trained with 1,191 images, it’s able to pick out Pokémon with an almost 97% accuracy.

In use, the system’s camera is pointed at a Pokémon model, then identifies these creatures on a nice looking user interface. While a Pi is implemented here, one could also run the program on your desktop/laptop as long as you have a webcam and are willing to lug it around. A quick demo of how the system works can be seen in the video below.

[h/t Raspberry Pi]