Image-based classification of plant genus and family for trained and untrained plant species.

摘要: Modern plant taxonomy reflects phylogenetic relationships among taxa based on proposed morphological and genetic similarities. However, taxonomical relation is not necessarily reflected by close overall resemblance, but rather commonality of very specific characters or similarity the molecular level. It an open research question to which extent relations within higher taxonomic levels such as genera families are shared visual constituting species. As a consequence, it even more questionable whether plants at these can be identified from images using machine learning techniques. Whereas previous studies automated identification focused species level, we investigated classification families. We used 1000 that representative for flora Western Europe. tested how accurate representation learned their in order identify included excluded learning. Using natural with random content, roughly 500 per required classification. The accuracy amounts 82.2% increases 85.9% 88.4% genus family Classifying training, significantly reduces 38.3% 38.7% Excluded well represented classified 67.8% 52.8% accuracy. Our results show indeed present levels. Most dominantly they preserved flowers leaves, enable state-of-the-art algorithms learn representations Given sufficient amount composition training data, this allows high increasing level facilitating process.