Image augmentation for object detection of grapevines

Abstract

Machine Learning methods are widely used for data analysis in various areas. In this work we use Neural Networks for image analysis in order to detect grape fruit clusters. A set of manually tagged images is built and a comparison is made between different data augmentation techniques in order to analyse the best way to expand the image set. The technique presented here obtained up to 13% better detection performance starting with only 100 images for training. The types of transformations and filters that worked the best for these images are discussed. In addition, training and detection times in five different hardware infrastructures, both CPU and GPUs, are briefly discussed.