Experimental Determination of CNN Hyper- 􀀳arameters for Tomato Disease Detection 􀁘sing Leaf Images

Abstract

Today, deep learning has become an emerging topic widely used in pattern recognition and classification problems. The design choice of the deep learning models entirely depends on who it's going to create. Still, it requires prior experience because identifying the best combination of parameters is a challenging task. So, the main objective of this study is to develop an accurate model for tomato disease classification while exploring the possible range of parameters that highly affects the performance of the Convolutional Neural Network (CNN). A simple CNN model was first built and train from scratch by using 22930 tomato leaf images collected from the Plant Village dataset in Kaggle. The model was tested for many cases by changing the values of a set of parameters while keeping other parameters constant. Finally, performance metrics were evaluated for every chosen parameter comparing with the base model. The highest prediction accuracy, training accuracy, and validation accuracy achieved from the study are 92%, 94%, and 92%, respectively. Rather than offering a guess, this study can, at most, give a definite answer that will assist new researchers in understanding how the accuracy and loss vary for every parameter within the area of tomato plant disease classification.