Mushroom Classification using Transfer Learning Techniques

Accidental intake of toxic mushrooms is still a very serious public health problem even in this modern age when mushrooms are consumed everywhere like food and medicine. Mycological identification has traditionally been dependent on expert knowledge derived from a visual appraisal of basic cap, gills, stem, and spore print characteristics; hence this is one of the reasons why, in many instances, edibles cannot be distinguished from poisonous mushrooms. These techniques can be error-prone and are often subjective. This study employs convolutional neural networks to facilitate the classification of mushrooms associated with image data for solving the problem. This study considers five classes that are often encountered from the dataset: "Mushrooms Classification - Common Genus's Images"from Kaggle. Transfer learning was based on ResNet50V2 architecture. Weights were pre-trained to enhance feature extraction and classification. Techniques of data augmentation were employed to enhance variability and robustness of the dataset. The model now possesses an accuracy of 81.41%, thereby proving its efficacy in discrimination among the species of mushrooms based on their observable features. This strategy has great potential to be successful with the creation of a scalable, economically feasible tool for safer mushroom foraging and consumption.