DeepMushroom is a fungal classficiation project using ResNet

iNaturalist.org is a citizen science website that allows people to upload images of unknown organisms for identification by other ecology enthusiasts. We collected images of identified fungi uploaded between 2017 and 2019 via the iNaturalist export tool. The images downloaded were then categorized by their species.

We use both python and golang scripts to download the images from iNaturalist.org. See here: main.go

Distribution

The data distribution is heavily skewed towards the few most common species. We remove the fungal species with less than 10 images for two reasons:

• If one species has less than 10 identification on iNaturalist.org, it indicates that it is not frequently occuring. Therefore, there is less value in the identification of such species.
• There is not enough data to effectively train the identification model. A species with less than 10 images will hurt the overall accuracy of our model

Since the images from MushroomExpert were identified by mycologists, we can use their images as a reliable validator to test the performance of our model.

Since we are in the very early stage of the experiment we built the model with the fast.ai library. The model will gradually switch to our own models utilizing pytorch as we progress.