FGVC8 based foliar diseases identification with use of multi-label classifiers

Abstract

Image recognition is achieving unprecedented success in many areas, while high requirements for the detection accuracy of objects and the characteristics of objects present in images remain difficult challenges. On the other hand, the continued popularity of Computer Vision and the constant improvement of the quality of image recognition algorithms are the best way to solve these problems. This approach became the impetus for writing this article. The article examines the effectiveness of using FGVC (FineGrained Visual Categorization) algorithms to obtain the best solution to the problem of recognizing apple leaf diseases. The problem is very complex, because the leaves with symptoms of diseases are often infected with many diseases at the same time, which makes the whole process difficult. After analyzing the possible approaches to solving the problem, it was decided to examine four methods: use the most primitive AI (Artificial Intelligence) approach, use classic methods of image classification, check deep neural network based on multilabel classification and use deep neural transfer learning multilabel classification with search cutoff optimization. The conducted research using a proven and representative set of training and test data made it possible to compare all the abovementioned methods and to draw conclusions about the accuracy of using each of the methods.