A COMPARATIVE ANALYSIS OF CONVOLUTIONAL NEURAL NETWORK MODELS FOR CHILI LEAF CURL VIRUS DETECTION

Abstract

A Convolutional Neural Network (CNN) model specifically designed for the detection of the chili leaf curl virus, a serious hazard to chili plants, is developed and evaluated in this work. The research commenced with the assembly of a dataset comprising 80 images each of healthy and virus-afflicted chili leaves, which were meticulously labeled and resized for consistency. To overcome the scarcity of the data, data augmentation techniques were used to generate more data for the model. Data was divided into a training, validation and test set and a CNN architecture with three convolutional layers, ReLU activation function and max-pooling was designed. The model was optimized with the Adam optimizer and loss function as a binary cross-entropy was monitored throughout the training process along with key indicators of model performance or KPIs (loss, accuracy, etc.). A total of eight variants of Convolutional Neural Network (CNN) models were developed for the detection of chili leaf curl virus, which poses a great threat to chili plants. The CNN architectures were systematically developed with varying numbers of dense layers, dropout rate, activation function and other parameters to evaluate the accuracy of the classification of Healthy and Virus infested chili leaves into the CNN model. All variants were rigorously trained and evaluated, with additional testing on independent sets to get unbiased performance metrics. The results show that the results obtained by the different variants vary in their accuracy, which ranges from 50% to 100%, and the loss values are taken into account according to the models' predictions. Some variants showed potential performance with high accuracy while others showed some weakness either by underfitting or overfitting, showcasing the difficulty of the classification task.  All things considered, this thorough investigation offers insightful information about how regularization strategies and architectural decisions affect CNN performance. The comprehensive documentation presented in this study offers transparency and clarity regarding the model's performance, paving the way for its potential application in practical scenarios involving chili leaf curl virus detection.