Context

I found some old Pokemon Cards while spring cleaning and decided to try building a deep learning classifier to spot fake cards. This is my first time training a deep learning model and thought that binary classification would be a good place to start. For collectors (like myself), it is rather easy to tell if a card is fake based on the back graphics alone, but not quite so for newcomers. Rare Pokemon cards are highly sought after by collectors, with some even reaching re-sale prices of hundreds of thousands of dollars. It would be important for collectors to tell a genuine card from fake card to avoid getting duped.

Content

Images of Pokemon cards were obtained by myself with a Canon 700D DSLR and 60 mm macro lens. All fake Pokemon cards were from the Generation 3 Series (Ex Ruby & Sapphire - Ex Power Keepers). Real Pokemon cards consist of an equal mix of Generation 1 (Base Set - Gym Challenge), 2 (Neo Genesis - Legendary Collection), 3 (Ex Ruby & Sapphire - Ex Power Keepers), 4 (Diamond & Pearl - HeartGold SoulSilver) , and 5 (Black & White - Dragon Exhalted) Series. Images were cropped and and resized to 256x256 with Irfanview. Images were then split into train (373 images; 250 real; 123 fake) and test (78 images; 50 real; 28 fake) sets.

Acknowledgements

Thanks to Kaggle mini-courses on computer vision for getting me started on this. My first trained CNN to classify the card images heavily inspired by Francesco Marazz's CNN for the Histopathologic Cancer Detection competition.

Inspiration

The dataset is still rather small, and might not encompass the wide range of counterfeit Pokemon cards on the market. Any model trained on this model might not be robust enough to detect all counterfeit cards. If anyone would like to help expand the dataset, you can contact me.