[2H1-OS-21] OS-21

[2H1-OS-21-03] Training Data Augmentation for Deep Learning by using 3DCG Fruit Model

〇Ryoma Takai¹, Kazuki Kobayashi¹ (1. Shinshu University)

Keywords:Deep Learning, agricultural field monitoring , fruits area detection

This paper proposes a method to detect apple fruits’ shapes from field monitoring images. The proposed method can detect the fruits’ shapes even if they are hidden by leaves or other fruits. We use a deep neural network with a way to create a large amounts of artificial field images as training data. Since the developed system uses 3D models of fruits, leaves, and trees and can easily retrieve their exact shape data of the fruits in the creation process. The proposed method also automatically creates the various 2D images as training data by changing the camera angles and positions in the 3D space. The created data includes annotations of hidden fruits by leaves, branches and other fruits.