One-vs-All Methodology based Cassava Leaf Disease Detection
Aryan Sagar Methil, Harsh Agrawal, Varadh Kaushik
Abstract
Cassava is the most important tropical root crop. Its starchy roots are a major source of dietary energy for more than 500 million people. It is known to be the highest producer of carbohydrates among staple crops. It is a very resistant crop however it is not immune to all viral and bacterial diseases. Cassava Mosaic Disease alone causes an annual loss of US$ 1.2 to 2.3 billion. It is a very concerning issue hence it is imperative to detect the disease at its early stages. Modern deep learning techniques can be very useful given the need for predicting diseases with high precision. We propose one-such deep learning-based solution involving Convolutional Neural Networks. Our work proposes a “One-vs-All” methodology for solving the task of classifying the four most prevalent types of cassava leaf diseases and healthy cassava leaves. For this we trained five different binary classifiers for the five different classes, each classifier used the Efficient Net B4 model as the base model followed by few fully connected layers. Our final multi-class classifier achieved an accuracy of 85.64% on highly skewed test data. Further, we deployed the model on Android using Android Studio, Java and XML for more accessibility of our classifier.