Field Plant Identification Through Indoor Imagery Using Image Augmentation and Object Detection Algorithms

Abstract

A growing global population and recent changes in climate make it increasingly necessary to incorporate recent advances in machine learning and robotics into agricultural practices. Plant detection is the problem of localizing and classifying all the plants within a given scene. Recent state-of-the-art object detection algorithms show promising results in detecting multiple objects and have great potential for outdoor plant detection. They require, however, a massive annotated plant dataset. It takes a great deal of time and expertise to manually annotate a plant dataset on such a large scale. We propose automating the plant annotation process in this thesis. Synthetic plant image datasets are generated in an outdoor setting by augmenting indoor images of plants that were captured and annotated automatically using a robotic camera system. We examine two different approaches: using image processing techniques to place plants on a soil background and using a generative adversarial model to generate fully synthetic outdoor datasets. We train two different plant detection algorithms on the synthetic datasets and evaluate the results on a manually-annotated outdoor dataset. Our best-performing dataset shows promising results for adoption in larger-scale automatic outdoor plant dataset annotation.