Image-based Classification of Honeybees
Stefan Schurischuster, Martin Kampel
Abstract
Behavioral analysis of honeybees is a key factor for keeping a healthy bee colony and therefore impacts our lives indirectly or even directly, due to their pollination of many plant species. Rapid growth of parasites like Varroa destructor is one of the main reasons for the elevated mortality of bee colonies. In this paper we are classifying bees into 'healthy' and 'infected' based on the presence of this parasitic mite. A camera facing the entrance of a beehive is acquiring the images used for a novel dataset, which is used to segment and detect Varroa destructor. We are comparing two classification methods based on AlexNet and ResNet as well as a semantic segmentation approach using DeepLabV3, with the latter achieving a per-class accuracy of minimum 90.8% with an overall f-score of 0.95. The evaluations are performed on a ground truth dataset with more than 13,000 manually labeled images of infected and healthy bees, which is made publicly available.