Synthetic bootstrapping of convolutional neural networks for semantic plant part segmentation.

• This paper investigates convolutional neural networks on large agricultural datasets. • Synthetic dataset bootstrapping and empirical dataset fine-tuning is researched. • Plant parts can be recognized on a per-pixel level. • We show only a small annotated empirical dataset of 30 images is required. • A large synthetic dataset for bootstrapping improves performance. A current bottleneck of state-of-the-art machine learning methods for image segmentation in agriculture, e.g. convolutional neural networks (CNNs), is the requirement of large manually annotated datasets on a per-pixel level. In this paper, we investigated how related synthetic images can be used to bootstrap CNNs for successful learning as compared to other learning strategies. We hypothesise that a small manually annotated empirical dataset is sufficient for fine-tuning a synthetically bootstrapped CNN. Furthermore we investigated (i) multiple deep learning architectures, (ii) the correlation between synthetic and empirical dataset size on part segmentation performance, (iii) the effect of post-processing using conditional random fields (CRF) and (iv) the generalisation performance on other related datasets. For this we have performed 7 experiments using the Capsicum annuum (bell or sweet pepper) dataset containing 50 empirical and 10,500 synthetic images with 7 pixel-level annotated part classes. Results confirmed our hypothesis that only 30 empirical images were required to obtain the highest performance on all 7 classes (mean IOU = 0.40) when a CNN was bootstrapped on related synthetic data. Furthermore we found optimal empirical performance when a VGG-16 network was modified to include à trous spatial pyramid pooling. Adding CRF only improved performance on the synthetic data. Training binary classifiers did not improve results. We have found a positive correlation between dataset size and performance. For the synthetic dataset, learning stabilises around 3000 images. Generalisation to other related datasets proved possible. [ABSTRACT FROM AUTHOR]