Wheat Ear Detection Using Neural Networks And Synthetically Generated Training Data

This paper investigates the usability of synthesized training data for the recognition of wheat ears using neural networks in the context of semantic image segmentation. For this purpose, we create detailed scenes of wheat fields consisting of 3D models with high-resolution textures and defined material properties. The generated scenes represent different states of maturity of the wheat plants. Afterwards, photo-realistic color images are synthesized, which also contain a binary image mask with the locations of the ear models. The resulting image pairs are then used as training data for two neural networks (U-Net and DeepLabv3+). To determine whether these data allow domain adaptation the trained networks are evaluated using real wheat field images. The IoU value of about 0.66 shows that information transfer from the synthesized images to real images is possible. Finally, we transfer the results to aerial images taken from an UAV. It is shown that the reduced resolution of these images significantly decreases the recognition rate.