Detection of Indonesian Tropical Fruit Plant Species Using Transfer Learning Method

Abstract

This proposed research aims to determine the best algorithm performance among VGG16, ResNet and MobileNet in the process of image classification of fruit plant species. Based on the results of research experiments, the best fruit plant image classification results are the results of implementation using ResNet. The accuracy of ResNet in the training stage, evaluation stage and testing stage was 94.65%, 89.28% and 87.72% respectively. The VGG16 model obtained the lowest accuracy, with results at the training stage of 3.36%, the validation stage of 3.36% and the testing stage of 3.33%. The low accuracy of VGG16 in the classification of fruit species can be attributed to several factors. One reason is the use of weak algorithms in some cases, which limits the ability of models to accurately classify fruits. In addition, training on a small number of datasets can also contribute to lower accuracy, as models may not be able to achieve reliability.