Automatic Flower Detection and Classification System Using a Light-Weight Convolutional Neural Network

Phenology describes the timing of life history events like flowering in plants and is a sensitive indicator of biological responses to climate change. However, recording plant phenology across space and time is a labour-intensive task. The creation of autonomous systems for in situ monitoring could greatly reduce expensive and time consuming manual human labour to both collect and analyze the data. One of the bottlenecks in creating such an autonomous system is computational complexity of the adopted Computer Vision methods. Deep Convolutional Neural Networks (CNNs) are state-of-the-art object detectors but can be very slow and computationally demanding. Light CNN topologies with only few layers which can achieve good performance are needed for lowering the processing power requirements. In this paper, we compare a light-weight CNN with two deeper CNNs on an object detection as well as image classification task on a dataset of Dryas flowers from Greenland.