Convolution Neural Networks Backbone model for Citrus Leaf Disease Detection

Abstract

In agriculture, Leaf disease inferred that the plant lacks elements, gets infected, or even the environment is not suitable and needs special treatment. Specific knowledge and experience were needed to classify the leaf disease. As a result, the Artificial Intelligence system to classify plant diseases was developed to help reduce the time needed and precision. The backbone model or the base model is the model that proved to be efficient in extracting the feature from the input images. This research aimed to find the backbone model that is suitable for citrus disease classification with localization. In this paper, Four backbone models chosen as a candidate were VGG16 [1], ResNet50V2 [2], DenseNet169 [3], and MobileNetV3 [4]. Both trainings from the scratch and transfer learning were used [5]–[8] to compare the model's compatibility and to detect Citrus leaf disease. The dataset [9] contains 596 images of diseased(canker, black spot, and greening) and healthy Citrus leaves with data augmentation. In this research, the model with transfer learning could achieve the best results in the most selected model. The models that have the best performance were VGG16, ResNet50V2, and DenseNet169, respectively. For the evaluation result of local collected data, The best model was VGG16 however the improvement was needed in the planed future work with the diseases detection with localization.