Generating artificial images of plant seedlings using generative adversarial networks

• Published in: Biosystems engineering Vol. 187; pp. 147 - 159
• Main Authors: Madsen, Simon L., Dyrmann, Mads, Jørgensen, Rasmus N., Karstoft, Henrik
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-11-2019
• Subjects: GAN; Image synthesis; Low inter-class and high intra-class variance; Plant classification; Plant seedlings; Plant seedlings; Low inter-class and high intra-class variance; Image synthesis; Plant classification; GAN
• ISSN: 1537-5110; 1537-5129
• DOI: 10.1016/j.biosystemseng.2019.09.005

Plants seedlings are a part of a domain with low inter-class and relatively high intra-class variance with respect to visual appearance. This paper presents an approach for generating artificial image samples of plant seedlings using generative adversarial networks (GAN) to alleviate for the lack of training data for deep learning systems in this domain. We show that it is possible to use GAN to produce samples that are visually distinct across nine different plants species and maintain a high amount variance within each species. The generated samples resemble the intended species with an average recognition accuracy of 58.9±9.2%, evaluated using a state-of-the-art classification model. The observed errors are related to samples representing species which are relatively anonymous at the dicotyledonous growth stage and to the model's incapability to reproduce small shape details. The artificial plant samples are also used for pretraining a classification model, which is finetuned using real data. The pretrained model achieves 62.0±5.3% accuracy on classifying real plant seedlings prior to any finetuning, thus providing a strong basis for further training. However, finetuning the pretrained models show no performance increase compared to models trained without finetuning, as both approaches are capable of achieving near perfect classification on the dataset applied in this work.
•Modelling plant seedlings from nine species using generative adversarial networks.•Model produces visually distinct samples for each species while still maintaining high variability within each class.•Artificial samples exhibit relative high resemblance to the intended species.•Applied in a transfer learning setup the artificial samples provide a strong basis for further training.