README.md

Mushroom Vision 4 🍄

This first project has the ambitious goal of building a deep learning model capable of classifying images of fungi into their species, providing information on the edibility or toxicity of each recognized species

The Mushroom Vision 4 app, currently running on HuggingFace, uses a DenseNet161 architecture to classify mushroom images into 4 classes.

The table below shows the species considered in this study (you can try these images by dragging and dropping them into the app).

	Species name	Common name	Photo	Edibility
01	Amanita Muscaria	Cocco del monte		Toxic
02	Amanita Vaginata	Amanita		Not edible
03	Boletus Edulis	Porcino		Great
04	Boletus Erythropus	Cappella Malefica		Toxic when raw

Feature extractors with 4 different architectures were tried:

1. AlexNet
2. DenseNet121
3. ResNet50
4. VGG16

The DenseNet121 performed better so all variants of this architecture available on torch were tested to identify the best performing:

1. DenseNet121
2. DenseNet161
3. DenseNet169
4. DenseNet201

The 161 form ouperformed the others as shown in the confusion matrices below.

The main goal was to minimize the classifications of poisonous mushrooms as edible, those that could potentially bring the most harm to potential users. In the validation set, no such errors are made while in the test set, probably due to the imbalance in the classes, misclassifications in this direction are present.

Despite this, the model offers very accurate performance as well as very high confidence levels on its predictions.

WARNING: This app is purely for demonstration purposes, has no application intent, and therefore whatever answers it gives are not to be considered reliable in any case. Only accurate mycological analyses can give reliable results about whether a mushroom belongs to a species and, therefore, whether it is edible. Wherever you go, whether to harvest or just to observe this wonder of nature always have respect for the places you visit, not polluting and not destroying nature.