Data Augmentation of Pseudo-Dense Images to Detect Morning Glory Regions in Soybean Fields

Morning glories prevent soybean growth and re- duce yields. However, early morning glories are small and diffi- cult to detect visually in the vast fields. Semantic segmentation effectively estimates the location of weeds using images captured by drones, but a large amount of image data is required to ensure high prediction accuracy, and an open dataset of plants with a wide variety of species is limited. In this research, we propose an image generation system of pseudo-dense morning glory using PGGAN to increase the volume of the dataset for training. The proposed pseudo images can configure the density of multiple morning glory using the distance. As a result of including these pseudo-densely morning glories, the F2 score of the estimation was 0.404.