Dropout Regularization to Overcome the Overfitting of the ResNet-50 CNN Algorithm in Oil Palm Leaf Disease Classification

The Residual Network-50 architecture is a way to overcome the problem of vanishing gradients in the Convolutional Neural Network algorithm. The weakness of the ResNet-50 architecture is overfitting, because the deeper training process causes the model to try to study the data thoroughly, including data that has noise. Overfitting the model requires a longer time in the training process and the test accuracy becomes lower. Overfitting can be overcome by using regularization techniques. This study proposed the use of dropout technique to overcome the problem of overfitting as a regularization on the ResNet-50 architecture. The dropout technique forces the model to learn different features during the training process. The ResNet-50 architectural CNN algorithm was used in the classification of oil palm leaf diseases with a total of 1692 data, namely 1128 training images (60%), 282 validation images (20%) and 282 test data (20%). The use of ResNet-50 without dropout obtained the best training accuracy results on epoch 10 of 97.23% and validation accuracy of 90.43%. The use of ResNet-50 with a dropout of 0.5 obtained the best training accuracy results on epoch 4 of 99.82%, validation accuracy of 94.14% and test accuracy of 98.60%