Automatic trait estimation in floriculture using computer vision and deep learning

Abstract

Registration of new varieties of ornamental flowers is an important process in protecting plant breeders' intellectual property as well as consumer rights. One of the first steps in the admission procedure for a new candidate variety is a consistent and thorough registration, leading to a description of a number of traits that should uniquely define each variety. Similar trait descriptions are used in other applications like distinctness, uniformity and stability testing (DUS testing). Typical traits relevant for ornamentals are flower color, color distribution and petal shape. For each species the set of traits will differ. This process is time consuming, susceptible to error, and depends on skilled expertise. In this work, we aim to increase the level of automation in this process by using computer vision to estimate/classify the selected traits from images of the flowers, considering real world data sets of roses and gerberas. Using standard deep learning architectures, accuracies of 35-99% have been obtained for selected traits.