The effect of improvement of datasets on accuracy achievement in deep learning: an example of disease detection in hops plant

Abstract

Plant diseases are a major threat to food safety and security. Preventing the loss of money and time is possible with early diagnosis of plant diseases. Recent advances in computer vision have led to successful methods for the early detection of plant diseases. In this research, images of downy mildew (Pseudoperonospora humuli) and powdery mildew (Podosphaera macularis) diseases of hops (Humulus lupulus - hops) plant were collected over the internet and classified with the most successful Convolutional Neural Network (CNN) model. In order to increase the performance of the CNN model, images that do not contribute to learning were removed from the datasets, and optimum datasets were created by adding new images that comply with the rules we determined. The model was trained with a small number of selected images and detected downy mildew and powdery mildew diseases of hops with high performance. In this study, certain rules were determined in the recognition of plant diseases, the collection of diseased leaf images and the creation of the data set. It has been shown that training datasets created by following these rules increase performance in learning.