Optimum Dataset Size for Ayurvedic Plant Leaf Recognition Using Convolution Neural Networks

Abstract

A Convolution Neural Network (CNN) model was developed by the authors for Ayurvedic Plant Leaf Recognition (APLR). This is a computer vision and multi-class classification problem. The motivation behind this problem was to create a plant leaf recognition system for locally available ayurvedic plants in and around the city of Visakhapatnam in the Andhra Pradesh State of India. The ayurvedic plant leaf image dataset was not readily available. The burden of leaf image dataset collection and creation fell upon the authors. The ayurvedic leaf image dataset creation was a tedious and time-consuming task. Due to the hot and humid environment in Visakhapatnam location, there was the problem of the quick withering of leaves upon plucking them. So, the leaves had to be photographed immediately in an outdoor location lest they withered and wrapped-. It is common knowledge that CNN requires a large training image dataset for creation of CNN model which can predict with high accuracy. There is no definitive guideline on the minimum number of required images per class in case of plant leaf recognition problem. This paper delves upon the relationship between the size of plant leaf image training dataset and prediction accuracy achieved using the CNN model proposed by the authors and trained using the corresponding plant leaf image dataset. The objective is to provide experimental results in this regard which will help in arriving at the optimum training dataset size to be used for APLR using CNN such that high accuracy is achieved.