Automatic Leaf Segmentation from Images Taken Under Uncontrolled Conditions Using Convolutional Neural Networks

Abstract

Automatic leaf segmentation from images taken in-field in uncontrolled conditions is a very important problem that has not been properly reviewed and that is crucial due to its possible use as a previous step in classification algorithms that can be used in agriculture applications. In this work, a CNN architecture (LinkNet) was trained to solve the isolated leaf segmentation problem under natural conditions. To do so, an open dataset has been modified and augmented, using rotations, shearing, and artificial illumination changes, in order to have a proper amount of imagery for training and validation. We have tested the CNN in two different datasets: The first belongs to the original open dataset that shares some visual characteristics with training and validation dataset. The second one contained its own imagery from a different set (images from different plants and with different illumination conditions) in order to evaluate the CNN model generalization. We obtained a mean Intersection Over Union (IoU) value of 0.90 for the first test and a 0.92 for the second one. An analysis of these results has been made and some problems regarding classification applications were commented.