On identifying leaves: A comparison of CNN with classical ML methods

Abstract

Convolution neural networks (CNNs) eliminate the need for feature extraction which is one of the most important and time-consuming part of traditional machine learning (ML) methods. However, the challenge of training a deep CNN model with a limited amount of training data still remains. Transfer learning and parameter fine-tuning have emerged as solutions to this problem. Following the recent trends, we address the task of visual identification of leaves in images by modifying a trained model on a similar problem. In particular, we show that a pre-trained CNN model on a large dataset (ImageNet) can be used to train a model from a small training set (ImageCLEF2013 Plant Identification). The resulting model outperforms the classical machine learning methods using local binary patterns (LBPs), a well utilized feature in the field.