Classification of tobacco leaf pests using VGG16 transfer learning

Some of the tobacco leaf pest attacks were only seen after the initial fermentation process. Tobacco leaves affected by pest attacks make the quality decline. Leaves affected by pests and diseases need to be separated from healthy leaves to maintain quality. Sorting is usually done manually allowing errors due to human-errors. In this study, we tried to classify the leaves affected by several types of pest attacks automatically. Convolutional Neural Network (CNN) is one of the latest classification methods proposed in this study using the famous VGG16 architecture. VGG16 training can last a long time if trained with random initialization of weights. For this reason, we selected initial weights by transfer learning to improve accuracy and speed up training time. Based on the results of training with 3-classes of the diseases using VGG16 and transfer learning, we obtained a very high accuracy. Some scenarios are tested based on a combination of the number of learnable parameters and types of the optimizer to get the best results. The result was that the proposed architecture was proven to be able to classify all training and validation data correctly. The dataset used was 1500 total images with 20% random cross-validation.