MushroomApp: a Mushroom Mobile App

Abstract

Background. Taking into account the mycological production of pine forests in Catalonia, more than 700 different species of mushrooms have been properly tagged and stored in a Data Base (DB). In this project we present MushroomApp. This App identifies mushrooms, by a simple image, from a corpus made up by the images of the DB. Supervised machine learning classifiers is an efficient mean for identifying mushrooms, and more specifically Artificial Neural Networks (ANN), so it was the one selected in this project. ANN models are created with Google Libray TensorFlow, positioned as the leading tool in the Deep Learning sector. Objective. The objective is to be able to create efficient ANN models using TensorFlow. In addition, we want to investigate a machine learning system to gradually improve our models. Methods. As there are many types of mushrooms, an important design decision was to mark the range of mushroms within the scope of the MushroomApp model. To implement the server we have used Python together with Django. The server is responsible for carrying out the operations of inserting new mushrooms and creating the TensorFlow models of the ANN. We will create these Models through Keras, a library that runs TensorFlow operations. The App is developed with Flutter to run the App on iOS and Android. Among its most important operations there are consulting the catalog, uploading images and making predictions. Results. The Precision, Recall and F-score of the ANN models for genus detection have been obtained with a corpus of 10,000 and 27,000 images. Mushroom detection performance has also been measured with 7,000 and 5,000 images. The best F-score obtained has been less than 0.5. Conclusions. The results obtained suggest an improvement and expansion of the corpus to increase the performance of the models obtained.