Understanding Deep Neural Networks for Regression in Leaf Counting

Abstract

Deep learning methods are constantly increasing in popularity and success across a wide range of computer vision applications. However, they are perceived as black boxes , due to the lack of an intuitive interpretation of their decision processes. We present a study aimed at understanding how Deep Neural Networks (DNN) reach a decision in regression tasks. This study focuses on deep learning approaches in the common plant phenotyping task of leaf counting. We employ Layerwise Relevance Propagation (LRP) and Guided Back Propagation to provide insight into which parts of the input contribute to intermediate layers and the output. We observe that the network largely disregards the background and focuses on the plant during training. More importantly, we found that the leaf blade edges are the most relevant part of the plant for the network model in the counting task. Results are evaluated using a VGG-16 deep neural network on the CVPPP 2017 Leaf Counting Challenge dataset.