Using Deep Convolutional Neural Network for oak acorn viability recognition based on color images of their sections

Abstract Convolutional Neural Networks (CNNs) are essential tools in many image recognition tasks. In this article we propose using a Deep Convolutional Neural Network for the task of visual assessment of the viability of mechanically scarified quercus robur l. seeds. Currently, this work is mainly performed by mechanical scarification followed by optical assessment by an employee. Here, we focus on acorn classification based on an innovative feature which is colour and intensity of image of sections of the seeds. We show that deep network accuracy (85%) is comparable or slightly higher than manual assessment of the viability of oak seeds. Also, we explore the impact of various image representations (colour, entropy, edges) as well as network architecture and its parameters on the classification results. We found that the image representation is a key factor in what the CNN is learning (topography of a pathological change or colour/intensity information only).