An Efficient Framework for the Classification Of Plant Leaf Diseases Using Optimized Machine Learning Algorithms

Abstract

This work proposes a leaf disease detection algorithm using Centroid Distance Neighbourhood (CDN) features and Genetic Algorithm based optimization. This method initially segment the disease affected region from the healthy region of the leaf. The disease affected region is applied for identifying the best feature points using Index Speeded Up Robust Feature (I-SURF) algorithm. The I-SURF feature points are used to select the features points. From a single I-SURF point, four features are extracted by forming a 5×5 neighbourhood across the SURF feature point. The feature extracted using CDN is optimized using Relative Fitness Genetic Algorithm (RF-GA) for best features that are able to classify the different diseases. newlineDuring testing phase the disease region is identified and features points are selected using the I-SURF points. The features are extracted using the CDN and the necessary feature that was optimized by RF-GA was sorted out as test features. The number of features reduces after the optimization process. These optimized features are used for training the classifier. The test features are classified from the trained feature using the K-Nearest Neighbour (KNN) algorithm to obtain the classification result. During training process the index value is calculated and it is used in the time of testing. So there is no need for performing RF-GA during testing process. The experimental result was evaluated on Plant Village dataset that are affected by various diseases such as Apple Scab, Bacterial Spot, Cercospora leaf spot, Common Rust, Black Rot and Healthy images. In this work, optimized features are used for training and testing is performed using the index values. This reduce the computational complexity as well as the average execution time. Most of the existing methods for computer aided detection of plant leaf diseases are based on highly complex machine learning algorithms that consumes more time for providing results. Lack of proper feature point identification, feature extraction and feature optimizati