Big bang–big crunch-CNN: an optimized approach towards rice crop protection and disease detection

Abstract

Convolutional neural networks (CNNs) are successfully used for solving image classification tasks. Manually designing efficient neural network architecture requires expertise in CNN and domain knowledge. In this paper, a novel approach is proposed for automatically evolving CNN architecture using big bang–big crunch (BB–BC) algorithm. The proposed approach can be used to automatically evolve CNN architecture and obtain tuned hyper-parameters for CNN. The CNN architecture is validated using the 5932 on-field rice leaf images infected with bacterial leaf blight, rice blast, brown leaf spot and tungro diseases. The proposed approach performed better than CNN evolved using genetic algorithm, SVM, KNN, decision tree, random forest with the test accuracy of 98.7%. The experimental results illustrate that the BB–BC CNN uses fewer trainable parameters compared to popular pre-trained CNN models like ResNet50, MobileNetV2, VGG16, VGG19, InceptionV3, etc.