A resource-efficient deep learning framework for pest and disease detection in plants

Abstract

The rapid growth of crop production has made early pest and disease detection particularly important in terms of efficiency and reduction of the use of pesticides. Conventional methods involving human experts are nonefficient for large crops. The recent advances in computer vision using deep learning can play an important role in increasing the accessibility and reliability. We apply a modified CNN based on MobileNetv2 to the task of pest and disease identification. We investigate various augmentation techniques to make the model more suitable for field conditions. However, the challenge here is that the performant deep learning models often require large computation and storage for mobile solutions. To overcome, we experiment with various optimization strategies to achieve a sufficiently lightweight implementation for Android OS. We use the dataset PlantVillage for disease classification which has 54K images and we achieve the accuracy of 96.3%. For pest classification we use the IP102 dataset having 75K images and achieve the state of the art accuracy of 64.8%.

Write-up of a machine-learning internship project at Wikilimo (May–June 2020).