<![CDATA[This paper discusses the application of Convolutional Neural Network (CNN) and Transfer Learning (TL) methods to improve the accuracy of dragon fruit classification. The application of the CNN method in real-time testing for classifying three types of dragon fruit only achieved an accuracy rate of 33.3%. Therefore, the CNN and TL methods using the Stochastic Gradient Descent (O-SGD) optimizer and the Root Mean Square Propagation (O-RMSProp) optimizer are proposed to improve the accuracy rate in classifying three types of dragon fruit: ripe, unripe, and rotten. The results of applying the CNN method with O-SGD at epoch 100 yielded an accuracy of 27.18%, val accuracy of 27.27%, loss of 1.407, and val loss of 1.405, while O-RMSProp at epoch 100 yielded an accuracy of 99.11%, val accuracy of 100%, loss of 0.073, and val loss of 0.058. Meanwhile, the application of the TL method with O-SGD at epoch 100 yielded an accuracy of 89.35%, val accuracy of 91.82%, loss of 0.462, and val loss of 0.443. TL with O-RMSProp at epoch 100 yielded an accuracy of 100%, val accuracy of 100%, loss of 0.002, and val loss of 0.003. The performance of the TL method with O-SGD and O-RMSProp is more accurate in classifying three types of dragon fruit compared to the CNN O-SGD and O-RMSProp models. This research contributes to improving the accuracy level of the CNN classification method to ±99-100%, and the application of this technology is an effort to enhance production quality and support smart agriculture in Banyuwangi Regency.]]>
