Hybridizing Convolution Neural Networks to Improve the Accuracy of Plant Leaf Disease Classification

Abstract

Plant leaf disease detection & classification is a complex image processing task, wherein proper algorithms are needed for segmentation, pre-processing, feature extraction and classification. Generally linear classification algorithms like support vector machines. (SVMs), k-nearest neighbour (kNN), Naive Bayes (NB), Random Forest (RF), etc. do not provide high precision for classification when applied to leaf disease classification. This is due to the fact, that the features which are evaluated during the segmentation and feature extraction phases are do not vary much in terms of values, but they vary in terms of patterns of occurrence. For example, leaf images which are taken for bacterial blight and Alterneria do not show significant changes in feature values, but they show major changes in feature patterns, which is generally neglected by these linear classifiers, and thus the accuracy reduces. In order to improve the accuracy of classification, we propose a hybrid convolutional neural network (CNN) in this paper, which combines multiple methods of segmentation & feature extraction with CNN in order to improve the accuracy of the system. The developed system shows 22% higher accuracy than the existing systems, and can adapt to any type of leaf images by moderate level of training.