Poisonous American Mushrooms Computer Vision Project

IMPORTANT: This project is for educational purposes only. The consumption of mushrooms can be incredibly dangerous. The creator of this project accepts no liability for injury sustained from consuming fungi.

The general purpose of my model is object detection of selected poisonous American mushroom species for use in real world foraging. The specific targeted use case is identification in the wild. Therefore, the images I have chosen to train my model are of mushrooms in their native habitat in order to best reflect this real-world use case. In addition to its practical application for foraging, this model allows foragers to better explore nature’s inherent mystery and beauty. With this particular approach, I am aware that the flaws of machine learning (ML) object detection, where a model cannot guarantee 100% accuracy, must be clarified from the outset.

The most specific subset of my intended audience is foragers, hikers, and mushroom enthusiasts who could use the model as a tool while foraging; it could also be used by environmental enthusiasts more broadly, ML professionals, and students. Ultimately, I foresee my use case as a tool comparable to iNaturalist, but targeted at foragers of mushrooms. Additionally, given that my 8 selected species are all native to the United States, my intended audience can be further clarified to be mushroom foragers within the United States.

I would never want anyone to use this model as the sole resource in determining whether or not a mushroom is poisonous; I aim for this to be a tool that potentially provides relevant insight and guidance, leading its users to do the necessary further research. Therefore, while I want my model to predict specific poisonous mushroom species in images, I want anyone using the model to be completely aware that any recommendation by my model must not be taken as objective truth, given the potential risk of poisoning and death.

The dataset for this model comprises 8 classes, each representing a different mushroom species, with approximately 100 images per class. The number of annotations varies across these species due to the different quantities of individual mushrooms present in each image.

The images and annotations were sourced from iNaturalist (for all 8 species), the Maryland Biodiversity Project (for the 5 species which could be found in Maryland), and Google Images (for 1 species), emphasizing a diverse and high-quality dataset. iNaturalist served as the majority source, offering thousands of example observations to choose from for 6 of my 8 species. Essentially, I used the Maryland Biodiversity Project and Google Images to complement iNaturalist. Overall, I was able to employ a balanced approach to my data collection, aiming to reduce bias by ensuring comparable diversity in the representation of each species while selecting only high-quality images to improve model accuracy. This diversity included different growth stages, orientations, angles, lighting conditions (with a focus on brightly lit), and environments to ensure the model could generalize well across various real-world scenarios. This methodology avoids bias to any particular appearance or setting.

This model was created for a class assignment in AI and Natural History at St. Mary’s College of Maryland.