Improved crossover based monarch butterfly optimization for tomato leaf disease classification using convolutional neural network

Abstract

To identify a plant disease accurately one should have a lot of experience and in-depth knowledge in a particular field. Identifying the plant diseases using manual intervention is often erroneous, time-consuming, and not a cost-effective option. Shallow Machine learning architectures were widely deployed for the automatic identification of the tomato leaf diseases, but their feature extraction process is highly time-consuming. Nowadays, the power of Deep Learning is been exploited by various researchers to identify the diseases present in plants. This paper utilizes a CNN approach to classify four different types of leaf diseases(bacterial spot, septoria leaf spot, late blight, and tomato mosaic virus) in a tomato plant without using any manual intervention. A dataset comprising of 6208 images of four classes of leaf diseases was acquired from the Plant Village database for classification. CNN is considered an effective option for solving a wide range of image processing tasks but its architecture is a little bit complex. To minimize this complexity and optimize the parameters present in the CNN, a binary solution encoding scheme is proposed using an Improved Crossover based Monarch Butterfly Optimization (ICRMBO) algorithm. This solution encoding technique implemented here eliminates the need for manual effort for designing the CNN architecture. Two convolutional architectures namely Vgg16 and Inception V3 were used in this work and they were optimized using the ICRMBO algorithm. The Inception V3 and Vgg16 architectures are widely deployed in this work to ease the training process, improve the generalization ability of the CNN network, and increase the classification accuracy. The overall test accuracies obtained for both the Vgg16 and Inception V3 was 99.98% and 99.94% when optimized with the ICRMBO algorithm. The proposed method augmented the performance in terms of high precision, sensitivity, and specificity values and obtained an overall classification accuracy of 99%. The significance and effectiveness of the proposed method over other methods are identified through the comparative analysis conducted.