An end-to-end deep learning system for hop classification.

Abstract

Automatic classification of plant species is a very

challenging and widely studied problem in the literature. Distinguishing different varieties within the same species is an even

more challenging task although less explored. Nevertheless, for some species distinguishing the varieties within the species can be of paramount importance. Hops, a plant widely used in beer production, has over 250 cataloged varieties. Although the varieties have similar appearances, their chemical components, which influence the aroma and flavor of the drink, are quite heterogeneous. Therefore, it is important for producers to distinguish which variety the plant belongs to in a simple manner. In this work, an end-to-end deep learning system is proposed to automate the task of hop classification. Five architectures are proposed and evaluated with an uncontrolled environment dataset that includes 12 varieties of hops on 1592 images, from three different cell phone cameras. The best architecture automatically detects the hop leaves on the image and performs the classification using the information of up to 10 leaves. The method achieved an accuracy of 95.69% with an inference time of 672ms. To reach such figures, state-of-the-art convolutional blocks were explored along with data augmentation techniques. Our results show that the system is robust and has a low computational cost.