Design and Evaluation of a Convolutional Neural Network for Banana Leaf Diseases Classification
Antonio Criollo, Miguel González-Mendoza, Eduardo Saavedra, G. Vargas
Abstract
In the agro-industry, plagues and diseases represent a great risk for the production because it affects the quality and quantity of the resulting product. The quick determination of these diseases, allows the production to be more efficient. In the present study, because of the current situation, it was chosen to take images from a free online repository. Convolutional neural network (CNN) was implemented to detect banana plant diseases from banana leaf RGB images. The development of 3 models was considered: The first one without regularization, the second one with Dropout and the last one with weight regularization. K-fold cross-validation method was used for hyperparameters tuning, Adam optimizer was used to accelerate neural network (NN) training. Also, metrics as accuracy and F1-Score were used to evaluate the NN performance. In the process, a total of 27 hyperparameter configurations were formed for each kind of model, from which the best configuration was chosen. Based on the results, it was concluded that CNN without any regularization technique was more effective for predicting banana plant diseases than the other models, despite having few data.