A Study on Applications of Transfer Learning Techniques to Plant Disease

Abstract

People must eat crops almost every day. Therefore, the health condition of all kinds of plants can be important. Incorrect judgment in the face of crop diseases can lead to the death of crops, which in turn affects the harvest of crops. Therefore, quick and correct judgment of crop diseases is very helpful to farmers. This project mainly compares the classification accuracy between convolutional neural networks with two different transfer learning techniques, MobileNetV2 and ResNet-50. The data used in this project are plants’ disease and health images. These data are collected from PlantVillage. By feeding the computer 8608 images of certain types of plant diseases, we trained deep convolutional neural networks to recognize plant species and classify plant diseases. However, due to the small number of images, the classification accuracy was quite low at first. So, we introduced two transfer learning techniques to improve the classification performance. We aim to find out whether transfer learning techniques are helpful in image classification and which one of the two complicated transfer learning techniques is better.