A deep learning generative model approach for image synthesis of plant leaves

What is it about?

The ability to generate realistic synthetic images of leaves has been explored since time in the field of computer graphics to create scenes with credible landscapes covered with plants, trees or meadows for use in computer games, virtual reality and entertainment industry. Current interest has been expanding to quantitative applications related to advanced approaches in botanics, plant breeding and agriculture. In this work we propose an approach, at the best of our knowledge not explored before, and, namely, we use implement Deep Learning techniques to automatically generate collections of synthetic images of plant leaves.

Featured Image

Photo by Rémi Walle on Unsplash

Photo by Rémi Walle on Unsplash

Why is it important?

We use the virtually unlimited synthetic images of leaves to enrich dataset of natural leaf images used for training Convolutional Neural Networks dedicated to smart agriculture applications. These models require indeed enormous amounts of annotated training images examples to avoid overfitting phenomena. Yet in real world applications annotated data are very often limited, especially for semantic segmentation tasks that typically require pixel-scale accuracy in manual labeling of training images.

Perspectives